Pricing trends

• Acqui-hire wave sweeps the AI tooling layer

  Sharpens

  37 companies

  The AI tooling and infrastructure layer is consolidating through acquisition — 26+ corpus companies have seen M&A or acqui-hire events in roughly two years, with LLM observability/gateways the hottest single segment (four exits in five months to May 2026: Langfuse, Helicone, Galileo, Portkey). But the DIRECTION of the wave inverted in mid-2026 and that is now the main claim. Through 2025 the acquirers were established platform companies (MongoDB, Elastic, Stripe, Kong, Zuora, CoreWeave, Nebius, Cisco, Palo Alto Networks, ClickHouse, Salesforce, Microsoft, Meta, Superhuman/Grammarly, Appier, Adyen) absorbing specialist tools. Since 2026-06-24 seven corpus companies have become acquirers themselves — Cursor→Continue, OpenAI→Northslope, Mercor→Deeptune, Harvey→Benchmark (its third in seven months), Vercel→Stakpak, Mistral AI→Emmi AI, Tempus→Personalis (~$1.5B) — including two frontier labs, which falsifies the earlier claim that no frontier lab was acquiring. The layer is now consolidating into its own winners as well as being absorbed from above. The deal-structure fork holds and gained its most explicit instance: Adyen's $335M acquisition of Orb runs an 'incubator' model that preserved operations, brand, multi-PSP support and the rate card, while acqui-hires (Clockwise, Robin, Rows, Humanloop, PlayHT, Continue) sunset or gate the product. A bounded sub-observation: un-publishing a once-public price precedes an exit by 8–18 months in 4 of 4 dated cases (Humanloop, Robin AI, Aleph Alpha, Inflection AI) but is 4-for-8 overall — Phind died 44 days after a raise with prices still up, OctoAI's sunset followed its acquisition by 36 days, and four companies un-published and kept trading. Start a clock, do not make a prediction.

  Unbabel Anyscale Gladia Tabnine Klaviyo OpenRouter Voyage AI OpenPipe OpenMeter Jina AI Togai Metronome Tavily Galileo AdCreative.ai Rows Clockwise m3ter Orb OpenAI Mercor Harvey Vercel Mistral AI Tempus HubSpot ScraperAPI Robin AI Langfuse Helicone Portkey Humanloop PlayHT Windsurf Reclaim.ai Windsurf (Cognition) Continue

  Reviewed August 2026 See the pattern

• Vendors withdraw public pricing as they mature or get acquired

  Sharpens

  33 companies

  A run of corpus vendors withdrew or hollowed out their public pricing through 2026 — Dropzone, Ada, Gladly, Clipdrop, OpenMeter, Unbabel, Robin, Artisan, 6sense, Codeium, Tavus, Inflection, Hyperline — most of them either repositioning up-market into regulated/enterprise buyers or being acquired, with the gate closing around the transition. Going gated is a maturation/acquisition tell, not a pricing-confidence signal. Four things sharpen the claim, updated at 380. First, the RATE has now fallen for three consecutive reviews: gated-or-sales-only is 108/380 (28%), against 103/353 (29%) and 30% at 338 — five more opaque companies in absolute terms, one point less in rate. The decomposition is the finding: the GATED bucket is frozen at exactly 48 companies across both reviews while the corpus grew by 27, so gated fell from 14% to 13% purely by dilution, and all five additional opaque companies landed in SALES-ONLY (55 → 60). Withdrawal is producing contact-sales walls, not login walls. Second, withdrawal is usually a STAGED ARC rather than an event — published ladder → free/entry tier killed and survivors repriced (often DOWNWARD) → numbers deleted — completed by six companies over 8-30 months, median 22, with Pixee the tell ($49 raise, $29 cut, deletion); a mid-arc discount reads as distress, not confidence. Third, and new, THE PRICE OF RECORD IS MIGRATING OFF THE VENDOR'S OWN SITE: Observe.AI's rates now live on AWS Marketplace while observe.ai/pricing 404s, and Poolside published its first-ever prices on OpenRouter and the Vercel AI Gateway while poolside.ai/pricing still 404s — so a dark /pricing page no longer implies an unknowable price. Fourth, there is now a level beyond gated: Inflection AI removed the pricing page, the quote request AND the contact-sales link, leaving no route to a price at all. The counter-current remains strong and increasingly comes from born-gated vendors publishing for the first time (Apollo's per-endpoint credit table, Freshworks' $49/100 bot sessions, Zilliz's first published entry prices, You.com's firm $1,200 replacing '>$2,000', Mercor's first dollar figure, Krisp's first self-serve door) — so transparency is a yo-yo, not a ratchet. SCOPE: this trend covers whole-page and TIER-level disappearances; the inverse shape, where a tier price stays published and the marginal/overage rate vanishes, belongs to the companion trend the-unit-rate-goes-dark. GitLab is the boundary case (seat prices behind 'Let's talk', all unit rates still public), and Flexprice is the false positive — its CTA moved while its $1000/mo list price stayed on the page. Two further distinctions land at 380. A PAGE can be withdrawn without the PRICE being withdrawn even inside one vendor: Groq (2026-08-11) 308-redirected groq.com/pricing to its homepage while every token, hour and character rate survived byte-identical in its own developer docs — only the three built-in agentic tools lost their public rate, because their docs pointed at the dead page. And opacity by BREAKAGE is not withdrawal: Higgsfield and InVideo both became legible on 2026-08-14 when client-side pricing tables that had failed since June finally rendered, revealing prices that had been public policy all along (and, at InVideo, no free tier where third-party sources had assumed one). Netlify (2026-08-04) marks the mildest genuine case at the other edge — deleting the 'starts at $500/month' floor caption from an already-quote-based Enterprise card is the smallest real number a vendor can remove.

  Dropzone AI Ada Gladly Clipdrop OpenMeter Unbabel Robin AI Sourcegraph Cody micro1 Lorikeet PhotoRoom Observe.AI Uniphore Phind Continue Flexprice Mercor Artisan 6sense Resemble AI Covariant Inflection AI Codeium Tavus GitLab Hyperline Groq Hyperbolic Moonshot AI Netlify 01.AI CrewAI m3ter

  Reviewed August 2026 See the pattern

• Voice APIs Converge on Per-Minute Billing

  Sharpens

  30 companies

  Voice and audio AI APIs have converged on media-minutes (per-minute or per-hour) as the dominant billing unit, displacing per-character and per-request alternatives. media-minutes is a billing unit at 48 of 380 corpus companies — count up 1 from 47/353, but SHARE DOWN from 13.3% to 12.6%, and now fifth-equal rather than clear fifth (seats 205, credits 149, tokens 74, api-calls 56, media-minutes 48, requests 47). The convergence is deepening inside the categories it owns rather than spreading into new ones. What is new at 380 is that the minute acquired DIMENSIONS: xAI (2026-08-04) became the first corpus vendor to publish two per-minute voice rates for two quality tiers of the same capability — grok-voice-think-fast-1.0 at $0.05/min and a new grok-voice-think-fast-2.0 at $0.08/min, a 60% premium — and Sarvam's Dubbing API (2026-08-15) bills per minute of source audio MULTIPLIED by target-language count, with an `editor_flow: true` request flag that exactly doubles the rate. The unit survives; the multipliers are new. It also keeps appearing inside non-voice vendors' cards: Glean's Model Hub added Deepgram Nova-3 at $0.0117/min next to per-million-token rows (2026-08-11), and LiveKit priced Gemini 3.7 Flash at $0.0029/min, exactly half Gemini 3.6 Flash's $0.0058 (2026-08-14). It spans real-time agent APIs (Vapi $0.05/min hosting, Retell $0.055/min voice infra metered to the second), enterprise telephony (PolyAI, Synthflow's three separate per-minute meters), WebRTC/agent infra (LiveKit agent-session minutes plus media minutes; Daily participant-minutes falling ~63% to $0.0015 past 50M min/mo), emotional voice (Hume EVI, deflating ~61% across three generations), STT platforms (Gladia, Speechmatics, AssemblyAI — all seven of AssemblyAI's priced products are per-hour), and legacy TTS vendors migrating off per-character for agent workloads. That migration was previously inferred; it is now documented as a single dated event: on 2026-07-29 Together AI unified all four serverless speech-to-text models at $0.0015 per audio minute, explicitly moving Parakeet TDT 0.6B v3 and Nemotron 3 ASR Streaming off per-1M-CHARACTER rates onto the minute and cutting Nemotron 3.5 ASR 67%. Two boundary conditions now qualify the claim. First, the minute is becoming the CAP even where it is not the meter — Lorikeet prices per voice resolution (1.50 / 1.20 credits) but footnoted that rate as applying only 'for resolutions up to 3 minutes long' (2026-07-21). Second, the all-in minute keeps shedding components: Bland AI removed telephony from its advertised per-minute rate, which now covers LLM + STT + TTS only (2026-07-21) — the minute survives as the unit while its contents are renegotiated (see voice-minute-unbundling). Genuine exceptions remain isolated: lmnt is per-character-only for batch TTS, Descript pools media hours at tier level rather than metering granularly, and Resemble prices detection per SECOND ($0.035/sec) rather than per minute. The structural driver is unchanged: AI voice is deployed in agent calls where wall-clock time, not text volume, is the cost driver.

  assemblyai speechmatics bland-ai elevenlabs cartesia deepgram tavus murf-ai rev-ai krisp synthesia twelve-labs wellsaid hedra fal-ai retell-ai vapi synthflow livekit daily hume-ai polyai gladia playht together-ai observe-ai lorikeet xai sarvam-ai glean

  Reviewed August 2026 See the pattern

• Free tiers moved in both directions at once — and layer only half explains it

  Sharpens

  29 companies

  Free-tier moves are running in opposite directions at the same time: of 23 dated moves, 16 narrowed a free tier and 7 widened one, and 22 of the 23 land inside a single 31-day window (2026-06-30 to 2026-07-30). The direction tilts with what the vendor sells — 13 of the 16 narrowing moves are seat-shaped applications and all 7 widening moves are developer-facing APIs or dev tools — but that tilt is our own coding of documented moves, not a vendor-stated rationale, and it is not a rule: Cerebras, Gladia and Novita are usage-metered developer APIs that narrowed anyway, and on the corpus's own `pricing_model_taxonomy` field the correlation weakens to 7 of 9 narrowing and 3 of 6 widening. The binding constraint is measurement: `has_free_tier` cannot see any of this, because a permanent recurring free tier and a 14-day trial both set it true.

  Poe Lokalise Qodo Speechmatics Bright Data Unstructured Artisan 6sense Freshworks Puzzle Cognition Cerebras Novita AI Gladia Photoroom Maxio Exa You.com Socket Hyperline Playground Gumloop Firecrawl SambaNova Mintlify ZenRows WellSaid Nomic Wispr Flow

  Reviewed August 2026 See the pattern

• Three usage-billing vendors win the buy decision — and Stripe Billing has pulled ahead

  Sharpens

  29 companies

  Among AI companies that buy (rather than build) their usage-billing layer, the vendor choice is concentrated but STRATIFIED BY LAYER AND VINTAGE, not a single three-horse race. Re-derived at 380 in_corpus companies (263 with a non-empty stack[] — the same 263 as at 353, so coverage fell from 74% to 69%): Stripe Billing 17, Orb 8, Zuora 6, Metronome 5, Chargebee 3, then a one-user tail (Lago, OpenMeter, m3ter, Togai, Hyperline, Paddle, RevenueCat, Maxio). CORRECTION: the 18 / 8 / 8 / 7 figures published at the 353 review were derived on this identical base and over-count name variants; the claim they supported survives the correction and strengthens, since Stripe Billing at 17 still outnumbers Orb (8) plus Metronome (5) combined at 13. Four distinct lanes: (1) Stripe Billing wins self-serve/PLG AI apps outright, and on 2026-07-22 STRIPE ITSELF DREW THE LANE BOUNDARY — its Billing page stopped marking usage-based billing as 'Included', replacing the checkmark with a Contact Metronome link and inserting Metronome as its own product card on stripe.com/pricing between Billing and Invoicing, labelled 'a Stripe product'. Billing stays 0.7% of billing volume for recurring subscriptions; basic metering survives via the Meters API to 100M events/month. Since Stripe acquired Metronome in January 2026, these two counts are no longer independent: 22 of the 263 classified-stack companies now buy billing from one owner. Meanwhile Orb added nobody — 8 at the 190-block read, 8 at 353, 8 at 380, across three reviews and a corpus that doubled; (2) Orb wins AI-infra vendors that already built their own meter and want only the invoicing/plan/credit layer (6 of its 8 users pair it with an in-house meter); (3) Metronome wins frontier labs and real-time inference infra where credit balances must resolve in-line with usage (anthropic, claude-code, replicate, together-ai, livekit); (4) Zuora — absent from the 190-block read and now named by 7 — wins incumbent, pre-AI-vintage sellers grafting usage onto an installed subscription base (canva, gitlab, zendesk-ai, yellow-ai). The lanes are not exclusive: Anthropic buys Metronome for metering AND Zuora for billing AND Stripe for payments, so the buy decision is per-layer, and 'which billing vendor' is the wrong question — 'which layer are you outsourcing' is the right one. Only 41 of 353 companies named any billing/metering vendor at the prior review, and 32 of those 41 named one of the top three, so concentration among disclosed buyers remains high (78%) — a figure not re-derived at 380 and carried forward. Two of the seven tail metering vendors are now owned by platform giants: Stripe acquired Metronome (Jan 2026) and Salesforce completed its acquisition of m3ter on 2026-07-21, with m3ter's quote-only /pricing page word-for-word unchanged. A third is rumoured — the WSJ reported on 2026-07-23 that Stripe is in TALKS (unconfirmed) to acquire OpenRouter for roughly $10B.

  Baseten Pinecone Vercel Relevance AI Replit AI Chroma fal.ai Fireworks AI Browserbase Intercom Langfuse Higgsfield Wispr Flow ElevenLabs Linear Anthropic Together AI LiveKit Replicate Manus Heidi Health invideo AI tl;dv Leonardo.Ai Bardeen Zuora Stripe Billing m3ter Yellow.ai

  Reviewed August 2026 See the pattern

• SKU-ladder velocity: vendors restructure the menu faster than they change the prices

  Weakens

  27 companies

  AI vendors restructure their plan ladders — adding, retiring, renaming, splitting and re-metering tiers — at a tempo that outpaces simple price changes, but the margin is narrowing fast and the claim is now WEAKENING. Across all 417 logged in_corpus change entries the corpus records 185 packaging events against 92 price-change events (2.01:1), down from 131:61 (2.15:1) at 353. In the fortnight to 2026-08-16 the ratio was 48 to 31 — 1.55:1, against 2.77:1 in the immediately preceding fortnight, a 44% collapse between consecutive cycles and the lowest figure ever recorded for this trend. Both components moved: packaging events fell 33% (72 to 48) while price-changes ROSE 19% (26 to 31), against a 25% decline in total activity. The published confirm condition (ratio above ~1.5:1 at n > 350 events) is still met on both cuts; the published kill condition (convergence toward 1:1) is one cycle away at this trajectory. What did NOT survive the re-test is the directional half. The original split — ladders expand for technical/self-serve buyers and collapse for enterprise — is contradicted by a single date: on 2026-07-22 Lago collapsed Business and Enterprise into one quote-only Premium package (consistent), Fathom ADDED a fifth, quote-only Enterprise tier and filled it by pulling SSO out of Team and custom retention out of Business (inconsistent — the enterprise end expanded, and by subtraction), and Inflection AI deleted its Enterprise offering entirely to a 404 (a third direction the split does not predict). Expansions in the window were not confined to self-serve (Fireworks split serverless into three serving paths; Observe.AI split its platform into five named modules; Hyperbolic into four compute surfaces; Composio added a $599 Business tier) and collapses were not confined to enterprise (Maxio 3 plans to 2; Mem0 5 to 4; LangSmith 7 meters to 2; Lovable 3 balances to 1; Freepik deleted Essential; Baseten trimmed to 8 models; Codeium's whole pricing surface retired into Devin). A second reason to downgrade appeared at 380: RESTRUCTURES ARE REVERTING, which inflates the packaging count without changing what a buyer can buy. Make split one plan into a Core/Pro/Teams ladder on 2026-06-11 and collapsed it back on 2026-08-11 (halving the entry credit step en route, with residual copy from the deleted ladder still live); Oxylabs flattened its Web Scraper grid on 2026-07-06, restored the two-tab layout on 2026-08-04 and flattened it again on 2026-08-11, with every per-1K rate byte-identical across all four states; Hyperline's flat $599 Launch tier lasted about a week before reverting to $299 + 0.7%. If a meaningful share of packaging events are round trips, the ratio has been measuring churn rather than menu evolution. The durable claims are therefore: (1) restructure tempo still exceeds repricing tempo, but at 1.55:1 in the latest fortnight rather than the ~2.7:1 previously claimed, and the gap is closing; (2) a ladder edit routinely carries a free-tier decision (Lokalise and Qodo deleted $0 plans; Flexprice re-metered Free to a $100K-billings cap; Gumloop, Freshworks, Photoroom and Cerebras all removed or converted free tiers inside restructures); and (3) restructures include forms no price diff can see — pure renames (Mercor relabelled three enterprise lines), entitlement subtraction at unchanged prices (Fathom), and meter normalization that doubles an effective rate (LangSmith). Direction of travel by buyer type is NOT predictable and should no longer be claimed.

  Tavus WellSaid Dust Runway Synthflow Frase AiSDR Sourcegraph Cody Windsurf Cursor Perplexity Ideogram HeyGen Galileo Qodo Vercel RunPod Weaviate Oxylabs AssemblyAI Flexprice Fathom Lago Fireworks AI Observe.AI Hyperbolic Mercor

  Reviewed August 2026 See the pattern

• Credits as the universal currency

  Sharpens

  26 companies

  AI vendors denominate *usage* in an internal 'credit' currency that floats across features — simplifying the invoice and decoupling the headline price from raw compute. The substitution test this trend was first scored on ('is credits the modal billing unit?') turned out to be the wrong test: at 380 companies seats (205) still lead credits (149), but 97 of the 149 credit vendors — 65% — also bill seats, so the two units were never rivals, and the ratio held flat at 0.727 through a 27-company intake of seat-heavy non-AI-native incumbents (0.729 at 353, 0.70 at 338). The credit is the usage layer stacked onto a seat base, and vendors keep bolting it onto previously seat-only products (Jasper 2026-07-22, Linear 2026-07-29). The live split is disclosure, not adoption: one branch pegs the credit to a published dollar rate (HeyGen $0.05/credit, LangSmith 1 LCU = $1.50, Hyperbolic 1:1 and never-expiring, Exa's dollar-denominated free credits), the other keeps it synthetic and unpriced (Jasper, Linear, n8n, Poe's compute points). The synthetic branch is where the price actually moves — Poe cut its $19.99 point allowance 34%, Ideogram quadrupled the credit burn per image on its newest model, and HeyGen raised Video Agent 50% — all inside a flat sticker price. Credits meter what you do, seats meter who does it, and the credit is the repricing lever.

  GitHub Copilot HeyGen Cursor (Anysphere) Recraft Descript Synthesia Codeium Cartesia Ideogram Lovable Creatify Bardeen Qodo Dust Zapier Jasper Linear LangChain Poe HubSpot Gumloop Exa AI Hyperbolic n8n ZenRows Framer

  Reviewed August 2026 See the pattern

• Cloud marketplaces as the enterprise billing rail — and now as the price-disclosure surface of last resort

  Sharpens

  25 companies

  Inference and compute vendors route enterprise billing through AWS / GCP / Azure Marketplaces so customers can spend down existing cloud commitments instead of signing a new vendor contract. 25 of 353 corpus companies now do it: anthropic, anyscale, assemblyai, baseten, cohere, dify, fireworks-ai, groq, hugging-face, lightning-ai, milvus, mistral-ai, modal, observe-ai, pinecone, qdrant, qodo, replicate, runpod, sambanova, together-ai, upstash, usageai, vantage, weaviate. It remains overwhelmingly an infra-layer move — but two boundary conditions identified at the 2026-07-30 review change how it should be read. First, it is no longer absent from the app layer: Observe.AI, a customer-service vertical, publishes a metered AWS Marketplace contract ($4.80/Minutes-per-Month unit, $12.00/Interactions unit, 12 months) with no seat dimension while its own pricing page 404s. Second, and more consequentially, the marketplace has become a PRICE-DISCLOSURE surface, not just a billing rail: for both Observe.AI and Usage AI the hyperscaler listing is now the ONLY place any price is published — Usage AI pulled its ~20% fee figure off its own site in mid-2026, leaving 20% of EC2 savings and 35% on RDS-class only in its AWS listing. A third extension is that the rail is generalizing past the big three, with upstash billing through the Vercel Marketplace. The trade-off buyers should price is unchanged and still almost never disclosed: hyperscalers take a marketplace cut, so the marketplace unit rate is typically above the direct rate, and Observe.AI's 24-month rates are exactly double its 12-month ones — a longer commitment with zero posted discount. Genuine counterexamples persist among seat-based verticals and consumer apps (harvey, glean, suno).

  Cohere Anyscale Modal Anthropic Baseten Fireworks AI Lightning AI RunPod Together AI Groq Replicate AssemblyAI Qdrant Dify Mistral AI Weaviate Pinecone Observe.AI Usage AI Hugging Face SambaNova Vantage Upstash Qodo Poolside

  Reviewed August 2026 See the pattern

• Sales-led AI buys the Salesforce stack; usage-metered AI builds the meter

  Holds

  25 companies

  A companys monetization stack forks on who it sells to. Sales-led, enterprise-contract AI companies buy the standard enterprise revenue stack — Salesforce CRM, NetSuite (or SAP) for ASC 606 rev-rec, and a CPQ/quote-to-cash layer (Salesforce CPQ, DealHub, or Conga) — because their pricing is custom-quoted and needs approval routing, not metering. Usage-metered, developer-facing companies instead build a metering pipeline and frequently name no CRM at all, because their pricing is computed from usage events, not negotiated. Across the 263 corpus companies with a classified monetization stack — the same 263 at a corpus of 380 as at 353, so coverage fell from 74% to 69% without a single company being added — the re-derived named-vendor counts are Salesforce 62, HubSpot 30, NetSuite 24, Salesforce CPQ 7, with a CPQ-CATEGORY entry at 30 companies, CRM at 114, metering at 91 and billing at 87. (CORRECTION: the 68 / 32 / 30 / 27 figures published at the 353 review were derived on this identical base and are an over-count from collapsing Salesforce name variants, not evidence of decline.) Category-level build-vs-buy, which is derivation-stable, is unambiguous wherever it is decided: metering 83% build of 53 decided, entitlements 88% build of 8, cost-finops 100% build of 8 — against billing 82% buy of 74, CRM 99% buy of 102, payments 99% buy of 94, rev-rec 98% buy of 42 and CPQ 94% buy of 18. Build the meter and the entitlement layer; buy everything with a contract attached to it. The fork's magnitude — and its direction — depend on how 'builds the meter' is operationalized, and both readings are published here because 74 companies file their in-house build under a generic `in-house` category rather than `metering`. Read broadly (any in-house-basis build component), 96 companies build and name no Salesforce against 47 that buy Salesforce with nothing built, 18 carrying both halves — the usage-built side roughly doubles the sales-led side. Read strictly (a stack entry whose category is literally `metering` with basis `in-house`), it is 39 usage-built against 64 sales-led with 4 both — the sales-led side leads. Same base, opposite headline; a single-number claim about this fork is not safe, which is itself the finding. The split still correlates with GTM motion rather than scale, and with category as cleanly as with motion: the horizontal platforms that ARE billing systems for their own customers build, while point-solution sales tools buy. HubSpot is the PLG-side CRM variant, so developer-first companies buy HubSpot+Gong instead of Salesforce while still building the meter. The premise is now contested from above, three times over: Salesforce completed its acquisition of the m3ter metering engine (2026-07-21), HubSpot shipped Revenue Hub as a priced CPQ/billing/payments product (2026-07-30), and the WSJ reported on 2026-07-23 that Stripe — the single most-named buy-side vendor in the corpus at 73 companies — is IN TALKS (unconfirmed, not announced or closed) to acquire OpenRouter for roughly $10B, which would put this trend's purest usage-built exemplar inside the payments rail the build side treats as a mere invoice endpoint. Buying the CRM no longer implies not buying a meter, and buying the payment rail may soon not either.

  Skydio Cohere CoreWeave Clay Cresta Harvey Scale AI Cursor Langfuse Firecrawl Tavus Lovable Kustomer Parloa Snorkel AI Observe.AI PolyAI Chroma Weaviate RunPod OpenRouter Salesforce HubSpot Stripe Chargebee

  Reviewed August 2026 See the pattern

• PLG Motion Locks In Public Pricing

  Holds

  24 companies

  A product-led on-ramp predicts a published price; a sales team predicts nothing; and a self-serve trial button without PLG predicts slightly worse than the corpus base. That three-way structure, computed across all 353 corpus companies, replaces the simpler "PLG mandates public pricing" claim. 210 of 224 PLG companies publish (94%), and the two PLG cells are where publication is effectively certain: plg+self-serve 38/39 (97%) and plg+sales-led+self-serve 169/180 (94%). Sales-led in any combination publishes 208/308 (68%) — indistinguishable from the 71% corpus base, so the presence of a sales team carries no information about transparency. The boundary condition is the intermediate cell: sales-led + self-serve WITHOUT PLG publishes 33/50 (66%), below the corpus base. A trial button bolted onto an enterprise motion does not force a price onto the page; only the product-led on-ramp does. And the mirror is the strongest single number here — of the 71 companies whose only motion is sales-led, 68 publish nothing (96% non-public), a 24× swing against the plg+self-serve cell. The mechanism is that PLG requires a buyer to complete a purchase decision alone. A trial can be gated behind a form and still be called self-serve; an activation funnel that must convert without a human cannot hide the price. Companies that lose the PLG on-ramp lose the public price with it — Artisan removed its free tier, credit estimator and self-serve checkout in the same 2026-07-14 change that removed its rate card.

  openai anthropic elevenlabs replicate fal-ai firecrawl exa-ai perplexity-ai cursor deepseek groq upstash modal turbopuffer tavily composio retell-ai vapi lovable n8n arize-ai metronome nomic chargebee

  Reviewed August 2026 See the pattern

• The allowance, not the price, is the repricing lever

  Sharpens

  24 companies

  Vendors are repricing through the included allowance rather than the sticker: 27 of 380 corpus companies moved what a plan includes — credits, tool calls, RAM ceilings, seats, storage, email caps, container concurrency, credit lifespan — while leaving every published dollar figure intact, 12 downward and 14 upward, with FLORA moving both ways in three days. The corpus-level shape backs it: packaging changes outnumber price changes 185 to 92 all-time (2.0:1), though the ratio narrowed this cycle to 48 to 31 (1.5:1) from 72 to 26 (2.8:1) last cycle, so the sticker has not been retired as a lever. The enabler is credit consolidation: once many per-feature meters collapse into a single currency, the burn table and the monthly grant become private price controls that no headline-price tracker reads, and the 4 corpus companies that consolidated and then devalued did so at a median 68 days. The allowance is not one variable but five — the monthly grant (Poe 1M to 660K points), the burn rate (HeyGen 20 to 30 credits/min), the capacity ceiling (Apify RAM, Modal's 1,000 to 5,000 containers, Instantly's 100K to 125K emails), the credit's lifespan (Kaiber turning forfeiture into rollover), and — new at 380 — the allowance's EXISTENCE, with Metronome deleting Starter's $100K/10M included volume outright on 2026-08-04 twenty-one days after publishing it, Speechmatics converting a recurring monthly free allowance into a one-time $100 grant, and FLORA switching a $12/$50/$100-per-seat included-usage promotion off and back on inside a fortnight. The boundary condition is that this is a two-directional lever, not a stealth-increase story — the single largest move in the corpus is still n8n growing an allowance 46-91x at an unchanged price, and 14 of the 27 moves run the buyer's way — so the claim is about where price now moves, not which way. The open risk is measurement: a corpus that tracks stickers is blind to both directions.

  Poe Ideogram HeyGen Composio Playground Vellum Apollo n8n Roboflow Apify Socket LlamaIndex Juicebox Firecrawl Exa Sarvam AI Kaiber Metronome Make FLORA Speechmatics CrewAI Modal Instantly

  Reviewed August 2026 See the pattern

• Vertical enterprise AI keeps pricing gated

  Holds

  24 companies

  Gating is not spreading — it is sorting. The corpus-wide gated-or-sales-only rate is 103/353 (29%), statistically flat against 100/338 (30%) seven weeks earlier and DOWN as a rate, so the 'rising opacity' framing is withdrawn. What the data supports instead is a stable 7.6× spread across product segments that predicts transparency far better than any 'enterprise' or 'vertical' label: customer-service 53% and fintech 53% gate at the top, then security 50%, vertical-SaaS 45%, ai-platforms 32%, horizontal-SaaS 28%, data-platform 20%, ai-coding 18%, developer-tools 14%, and infra-cloud just 7%. The mechanism is what the buyer can walk away to. Where a comparable public alternative exists one layer down the stack, gating costs the deal — infra-cloud vendors sell GPU-hours by the second against a public spot market and cannot gate at all. Where value depends on the buyer's headcount, regulatory posture or data estate, and no per-unit comparison is possible, gating is nearly free — which is why contact-center AI, clinical AI, legal AI and billing infrastructure all cluster above 45%. The direction of individual moves remains two-way and roughly balanced: four corpus companies gated further in this window (Artisan reverting to quote-only for the second time, Inflection deleting its enterprise surface entirely, Flexprice moving Scale to Contact Us, Hyperline dropping its published take-rate) against at least one de-gating (Apollo publishing per-endpoint credit rates). Expect the corpus rate to stay near 29% and the segment spread to stay wide.

  Harvey Glean Jasper Spellbook Lago Metronome Orb Parloa PolyAI Uniphore Observe.AI Mercor Robin AI Legora Nabla Suki AI Artisan Inflection AI Flexprice Hyperline m3ter Groq Moonshot AI Netlify

  Reviewed August 2026 See the pattern

• Self-serve list-price inflation

  Sharpens

  23 companies

  Application-layer AI vendors on self-serve plans are raising headline list prices in mid-2026, often by double or triple digits, even as the inference infrastructure beneath them keeps deflating. The driver is margin recovery: as the early land-grab phase ends and free GPU-price tailwinds get competed away at the infra layer, app vendors that had under-priced to acquire users are now repricing toward sustainable unit economics — frequently bundling the hike with a meter change (more included allowance, a credit pool, or a model toggle) to blunt churn. As of 2026-07-30 two refinements matter more than the list-price cases. First, most of the increase now arrives through mechanics that leave no list-price diff at all: SHRINKFLATION, where the sticker holds and the entitlement shrinks (Poe cut its $19.99 allowance 34%; Composio cut Pro's included tool calls 75% at an unchanged $29; Fathom pulled SSO out of Team; LangSmith moved Email Support off the $39 Plus plan; Arize deleted its published overage rates; Cartesia removed its annual toggle for a ~25% effective rise; Freepik cut its annual discount from ~40% to 25%), and FLOOR INFLATION, where an inference vendor retires the cheap end of its catalog rather than repricing it (Baseten's 11→8 SKU cut lifted its cheapest cache-input rate $0.06→$0.12 with no model repriced; Groq dropped its two cheapest LLMs; W&B took its floor model 14x/28x, lifting the catalog floor $0.01→$0.05). Second, the app-vs-infra 'scissors' has broken: the infra layer raised in the same window (Hyperbolic's H100 SXM +93% to $2.89/hr, Together AI's reserved H100 tenors up for the first time since June, DeepInfra's second raise in 15 days) while app-layer vendors cut (Juicebox's Starter -14% with doubled credits; Intercom's Essential discount revived on both cadences, reversing the promo-expiry raise logged in June — and now served as an allocated A/B test tagged data-experiment='team-self-serve-essential-price-reduction'). Tavus added and removed the same introductory promo inside eight days. The pattern is selective repricing on both layers, executed through entitlements and catalog composition more often than through list prices. The open risk is unchanged: coordinated self-serve hikes invite churn in low-switching-cost categories (email, creative, outbound).

  Shortwave Creatify lemlist Dust Lokalise Tavus Mistral AI Intercom Poe Zendesk Zhipu AI VEED Freepik Cartesia Hyperline Stripe Billing Fireworks AI Hyperbolic fal DeepSeek Make Mintlify SambaNova

  Reviewed August 2026 See the pattern

• Spend guardrails arrive after the bill-shock backlash

  Sharpens

  22 companies

  Spend governance is now a packaging feature across credit- and usage-metered vendors, and it has stratified into three generations of increasing sophistication. Gen 1 is reactive: an apology, a refund, and an advance-notice policy after the incident (Cursor's 30-day commitment, July 2025). Gen 2 is a structural cap that makes overspend impossible instead of refundable — prepaid wallets (Gladia 2026-07-22, Cerebras $5 trial 2026-07-21, Hyperbolic never-expiring credits with Auto Top-Up 2026-07-21), flexible spending credits (Vercel's $20), hard daily allotments (Lovable, Manus), concurrency ceilings (Vapi's 10 lines then $10/line), and non-purchasable credits that cannot be topped up at all (n8n 2026-07-22). Gen 3, appearing for the first time in mid-2026, is predictive disclosure: publishing the DISTRIBUTION of consumption so a buyer can forecast variance rather than just bound it — Glean's p50/p90 rate card across 13 capabilities (2026-07-22) and Clay's 75th-percentile withhold-then-refund (2026-06-02). Gen 3 matters because caps solve the wrong problem: a buyer whose Glean Agent Run costs 7 FlexCredits at p50 and 114 at p90 needs the spread, not a ceiling. Two failure modes recur. Guardrails ship WITH generosity cuts (Gladia traded 10 recurring free hours a month for a one-time 50 EUR grant in the same change that added the wallet), and guardrails can reduce legibility rather than increase it (Lovable merged dollar-denominated Cloud and AI balances into abstract credits on 2026-07-21). And the anti-pattern is now documented: Braintrust measures retention storage at ingestion, so deleting data or shortening the window does not lower the bill — a meter with no buyer-side lever at all.

  Cursor (Anysphere) Vercel GitHub Copilot Firecrawl ElevenLabs Manus Runway Augment Code Clay Lovable Vapi Synthflow Glean Gladia Hyperbolic n8n Cerebras Zapier Fireworks AI Diffbot Modal Speechmatics

  Reviewed August 2026 See the pattern

• Token prices fall as access tiers rise

  Sharpens

  22 companies

  Per-token API prices fall relentlessly with each model generation and via caching/batch discounts, even as consumer subscription ceilings climb — compute deflates while the price of human-facing access inflates. Deflation now bends even long-held anchors (Claude Sonnet 5's introductory $2/$10 undercut the four-generation $3/$15 Sonnet band) and still holds at FIXED capability: Anthropic shipped Opus 5 at exactly Opus 4.8's $5/$25, and Hugging Face's new AWS H100 at $4.50/hr undercuts its own GCP H100 at $10/hr. But it is not monotonic, and as of 2026-07-30 the exceptions have a repeatable mechanism rather than being anecdotes: years of cuts → CATALOG PRUNE → selective raise on the cheapest or newest SKU. Four vendors complete that arc (DeepInfra, Together AI, Novita AI, Moonshot AI) and 9 raises landed across 7 vendors in 50 days. The prune is the load-bearing middle stage — it lifts the effective floor with no visible hike on any headline model: Baseten's 11→8 SKU cut raised its cheapest cache-input rate from $0.06 to $0.12 while repricing nothing, Groq's 8→6 LLM cut removed its two cheapest models with every retained price stable, and Novita delisted 49 of 221 models in 8 days while taking one image SKU 10x ($0.036 → $0.36). Separately the frontier ceiling re-inflates by FIVE distinct routes: a higher flagship (Moonshot's Kimi K3 at $3/$15, 3.2x/3.75x Kimi K2.6; xAI's grok-4.5 above the still-live grok-4.3), a repriced EVERYDAY tier (OpenAI's GPT-5.6 Luna at $1/$6 vs the outgoing GPT-5.4 mini's $0.75/$4.50; Google's Gemini 3.5 Flash-Lite at $0.30/$2.50 above Gemini 3.1 Flash-Lite's $0.25/$1.50), an agent surcharge (Google's AlphaEvolve at 2x base, 3x all-in), a routing/quality premium (Fireworks' flat 10% for US-only Serverless, its 1.5x region-restricted dedicated deployments from 2026-08-11, and its Fast paths at roughly 2x Standard), and — new at 380 — TIME OF DAY. DeepSeek, whose V2 launch at $0.14/1M set the industry's floor expectations in May 2024, split its API into peak (01:00-04:00 and 06:00-10:00 UTC) and off-peak windows effective 2026-08-16, taking V4-Flash cache-miss input from a flat $0.14 to $0.22/$0.44, output from $0.28 to $0.66/$1.32, and cache-hit input from $0.0028 to $0.007/$0.014 (over 12x on V4-Pro), with no grace period or legacy-rate opt-out. The first four routes price a DIFFERENT product; time-of-use prices the same product differently by the clock. What deflates is the price of a fixed capability level — and it is deflating faster than ever, with GPT-5.6 Luna cut roughly 80% within twelve days of launch and the cut propagating identically through five corpus resellers in eight days. What inflates is the floor of the catalog, the price of the best, the price of the agent on top, the price of where it runs, and now the price of when you run it.

  Anthropic Novita AI OpenAI Google DeepSeek Hume AI Together AI Vercel Speechmatics Lambda Labs You.com Character.ai Mistral AI xAI DeepInfra RunPod Baseten Groq Weights & Biases Moonshot AI Hugging Face Fireworks AI

  Reviewed August 2026 See the pattern

• Infrastructure-Layer AI Vendors Standardize on Commitment Pricing

  Holds

  21 companies

  Infrastructure-layer AI vendors pair public PAYG pricing with commitment or reserved-capacity discounts at 28 of 42 infra-cloud companies (67%), against a corpus baseline of 154/353 (44%) — a 23-point lift that is the widest of any product segment. But the absolute rate is falling as the segment broadens (79% at 19 companies, 57% at 21, 67% at 42), so "almost universally" is withdrawn. The economic logic is unchanged and bilateral: providers with fixed silicon costs cannot absorb demand uncertainty, and buyers want throughput or latency SLAs, so committed use at 20-55% off PAYG lets both sides plan capacity. Lambda Labs remains the widest spread in the corpus — B200 from $2.99/hr on a multi-year commitment against $6.69 on-demand, roughly 55% off. What is new is that the commitment is migrating in two directions at once. DOWN-market: Hyperbolic made Reserved capacity self-serve on 2026-07-21 in terms from one week to one month, prepaid in full, no early termination — a commitment a developer buys with a card — in the same window that it nearly doubled on-demand H100 SXM from $1.50 to $2.89/GPU/hr, making the commit the only route back to the old price. UP-market: RunPod published a dedicated enterprise surface on 2026-07-22 naming reserved baseline plus usage burst on one post-paid invoice, with clusters to 10,000 reserved GPUs. And the commitment discount is not monotonically spreading. Bright Data deleted the two deepest rungs of its committed ladder on 2026-07-14 ($1.1/1k at $999/mo and $1.0/1k at $1,999/mo both gone, leaving one $499/mo Scale tier) and added a recurring free tier instead; CoreWeave removed its reserved-storage discount on 2026-07-21. Where capacity is abundant, vendors are choosing acquisition over utilisation hedging.

  anyscale baseten browserbase cerebras deepinfra e2b fireworks-ai groq lightning-ai modal replicate runpod together-ai turbopuffer vast-ai lambda-labs weaviate pinecone milvus bright-data hyperbolic

  Reviewed August 2026 See the pattern

• Prompt Caching as a Structural Pricing Feature

  Sharpens

  21 companies

  Prompt and context caching is no longer an optional feature — it is a required column on any token rate card, and the DEPTH of the discount has become the competitive variable. 17 of the 74 token-metered corpus companies publish an explicit cached-input rate (23%, against 16 of 73 at 353), and the frontier of that group has converged on exactly 10× (90% off): Anthropic Opus 5 at $0.50 read vs $5 input, Moonshot Kimi K3 at $0.30 vs $3.00, SambaNova MiniMax-M2.7 at $0.06 vs $0.60, and Baseten's GLM-5.2 after a 46% cached-only cut on 2026-07-29 ($0.26 → $0.14 against $1.40 input, with output untouched). Vendors move the cached line while holding headline rates — xAI cut grok-4.5 cached input 40% ($0.50 → $0.30) on 2026-07-21 with $2.00/$6.00 untouched — because it buys the repeat-context agent workload without repricing the shop window. The mechanic is NOT one-directional, and three cases at 380 prove it. The cache is now also a price-INCREASE vehicle: DeepSeek's peak/off-peak split (effective 2026-08-16) raised cache-hit input more than any other line, from $0.0028/1M flat to $0.007 off-peak and $0.014 peak on V4-Flash and from $0.003625 to $0.022/$0.044 on V4-Pro (over 12×), collapsing the corpus's deepest cache ratios from 50× and 120× down to about 31× and 30×. The cached line can move AGAINST its own headline — Weights & Biases cut DeepSeek V4-Pro input 34% and output 27% on 2026-08-11 while raising cached input 43% ($0.14 → $0.20), shrinking the discount from 12.4× to 5.75× on a SKU that got cheaper. And a published cached rate can grant nothing: Perplexity's Gateway API lists nemotron-3-ultra at $0.25 input and $0.25 cache-read — a 1× discount on the same card that gives kimi-k3 and glm-5.2 a clean 10×. Four sub-mechanics the original framing missed: caching is TWO-SIDED (Anthropic charges 1.25× base input to write a 5-minute cache and 2× for a one-hour cache, so a low hit rate makes caching more expensive than not caching); the cached line inherits every other multiplier on the card — serving path, context band, and now the CLOCK, since DeepSeek's peak window doubles the cache-hit rate along with everything else; the cache is becoming a metered RESOURCE, not just a discount (Zhipu lists a distinct "Cached Input Storage" line marked "Limited-time Free" on every model that supports it); and at-cost RESELLERS cannot use the cache as a lever at all — Cursor, Augment Code, v0/Vercel and Glean each cut GPT-5.6 Luna's cache write $1.25 → $0.25 and cache read $0.10 → $0.02 in exact lockstep with its 80% headline cut, because pass-through pricing forbids an independent cached line. The wedge is unchanged: stable-prefix workloads get 10× cheaper input while one-shot and diverse-prompt workloads pay full price, and the cache investment is provider-specific, so switching costs rise with hit rate — Poolside's OpenRouter trailing average shows a 97.7% hit rate turning a $0.09/$0.18 list into $0.011/$0.179 effective. But the cache is also the line a vendor raises first when it needs margin, and it can be raised without touching the headline. The laggards are still conspicuous — Mistral and Cohere have published no cached-input rate 24 months after Anthropic shipped one.

  anthropic openai google-gemini deepseek together-ai groq fireworks-ai baseten replicate deepinfra xai moonshot-ai minimax sambanova novita-ai zhipu-ai braintrust poolside weights-biases perplexity-ai cursor

  Reviewed August 2026 See the pattern

• Build the meter, buy the biller: the AI-infra monetization split

  Sharpens

  19 companies

  Usage-metered AI-infrastructure companies treat the metering pipeline as core IP they build in-house, but treat invoicing and payments as a commodity they buy. The recurring shape is an in-house meter (events → aggregation → rating) that ships usage out-of-band to a third-party billing system of record for invoicing, plans and payment. Recomputed at 353 companies (263 with a classified stack), the split is sharper than previously published: metering resolves in-house 43 / vendor 9 — a 4.8:1 build ratio, not the ~2:1 carried from the stale 190-block snapshot — while billing-plus-payments resolves vendor 113 / in-house 12, a 9.4:1 buy ratio, and payments specifically is in-house 1 / vendor 93. 19 companies disclose exactly the paired build-meter / buy-biller split against only 9 running the fully-bought inverse. (On a broader read that counts in-house builds filed under a generic category rather than `metering`, the build share is 90/99 = 91% against a 49% billing build share and the paired set is 22; direction identical, magnitude definition-sensitive.) Two refinements as of 2026-07-30. First, an evidence GAP rather than a finding: of the ~90 companies with an in-house meter exactly 1 (Kustomer) is outcome-based, against a 7% corpus base — the metering basis of the 24 outcome-priced vendors is almost entirely `unknown`, so the corpus has no sourced evidence of how vendors meter the resolutions they bill for. Second, the 'biller' being rented is consolidating into the platform owners: Stripe routed usage-based billing out of Billing into its own Metronome with a published boundary at 100M events/month on the Meters API, Salesforce completed m3ter, and HubSpot turned its in-house quote-to-cash into a sold product (Revenue Hub, 2.9% + $0.50 per card transaction, 0.75% on sellers who keep Stripe). 98 corpus companies name Stripe in some form and 10 name a cloud marketplace as the billing system of record. The counter-patterns remain the fully-bought stack (the vendor meters too) and the fully-built stack (no vendor at all) — but 'fully built' is now a route to becoming a vendor.

  Chroma Weaviate RunPod Baseten Cursor Perplexity AI Vercel Langfuse Pinecone Relevance AI Lovable Suno Synthesia Dust Anthropic Stripe Billing Salesforce HubSpot Chargebee

  Reviewed August 2026 See the pattern

• AI companies are staffing a cost-FinOps role to defend the margin under the price

  Holds

  18 companies

  A distinct margin-discipline cohort exists, and it takes two forms rather than one. Where a company's price is fixed (a flat seat, a per-clinician license, a bundled AI feature, a per-unit credit) while its AI compute cost per unit is not, it either STAFFS a cost-FinOps / strategic-finance / infrastructure-cost role or BUILDS an in-house inference-cost engine, so the spread it keeps becomes an owned function rather than a residual of the cloud bill. Of the 307 corpus companies carrying a monetization_signals block, 12 disclose the staffed role (6sense, abridge, apptronik, canva, langchain, lightning-ai, roboflow, runpod, sugarcrm, tavily, zapier, zoominfo) and 8 disclose a built cost engine as a cost-finops stack entry (bentoml, deepinfra, deepseek, groq, hubspot, modal, sugarcrm, usageai) — 19 distinct companies, 6.2% of blocks, with sugarcrm the only company on both sides. cost-finops is the smallest category on both axes: 8 of the 263 companies with a non-empty stack (payments has 105, metering 89) and 12 of the 176 with hiring data (revops has 127, customer-success 118). The build side clusters on inference vendors whose gross margin IS fleet utilization; the staffed side has broadened decisively into incumbent seat-priced SaaS that bundled AI into an existing price (6sense 2026-06-11, sugarcrm 2026-06-29, zoominfo 2026-06-30 — all three added in the five weeks to 2026-07-30), where margin defense must happen entirely on the cost side because there is no usage price to raise. This is the cost-side mirror of the value-metric/outcome-pricing thesis: outcome pricing engineers the value side of the meter, cost-FinOps engineers the COGS side. It is still a cohort, not a norm — assemblyai (pure per-hour audio, in-house meter and LLM gateway) and firecrawl (first-party credit ledger; on 2026-07-30 it bundled +1,000 free search credits into every tier) carry the same GPU/COGS exposure and disclose neither the role nor the engine.

  Canva Zapier LangChain Lightning AI Abridge Tavily Roboflow Apptronik Modal ZoomInfo 6sense SugarCRM DeepInfra DeepSeek Groq BentoML HubSpot AI21 Labs

  Reviewed August 2026 See the pattern

• Outcome-based pricing

  Holds

  18 companies

  Charging per resolved outcome or completed task — not per seat, message or token — is a category standard inside AI customer support and a rarity everywhere else. The vertical concentration is now measurable: 14 of the 19 `customer-service` companies bill resolutions (74%) against a 4% corpus base, and of the 15 companies billing `resolutions`, 14 are also tagged outcome-based, 14 are in customer-service, and 14 have no free tier (against a 61% corpus base). Corpus-wide the unit is not spreading — outcome-based tagging held at exactly 24 of 353 companies (7%) through a 15-company corpus expansion and a 140-change window with no new adopter — so what is changing is the unit's SHAPE, not its reach. Four expansions are now documented: the time horizon (Sierra's Horizon, 2026-07-16, bills a business goal pursued over days-to-months rather than one conversation), a duration boundary (Lorikeet caps its voice-resolution credit rate at 3 minutes, above which the outcome is quoted), disclosure via third-party marketplace (Observe.AI publishes $12.00/Interactions-unit and $4.80/Minutes-unit on AWS Marketplace while its own /pricing 404s), and outcome-CONTINGENT pricing reaching infrastructure (RunPod Overdrive's pay-nothing-if-we-miss guarantee; Usage AI's 20%/35% share of realised cloud savings). The enabler remains attributable, countable work product. The risk has shifted: because the unit is table stakes, price competition has moved into the DEFINITION of an outcome — and outcome vendors are the corpus's most gated cohort (16/24 = 67% don't publish, vs a 29% base) with almost no free tier to test the definition on. Directional movement is now two-way at the dollar level: HubSpot held $0.50/resolution through a full brand deletion while Gorgias raised its overage 67% to $1.50.

  Intercom Rox Gorgias Ada Lorikeet Kustomer Parloa Pixee Digits Recursion Mercor Micro1 Observe.AI HubSpot Sierra RunPod Usage AI Yellow.ai

  Reviewed August 2026 See the pattern

• The $200 prosumer ceiling

  Sharpens

  17 companies

  The ~$200/month top tier is a corpus-wide Schelling point rather than a consumer-AI phenomenon, present at 17 of 380 companies (4.5%, unchanged from 16/353). The NUMBER converges; the NAME does not, and the 2026-07-30 claim that it did is now wrong — 'Max' appeared in-window at Higgsfield ($79/mo) and InVideo ($85/mo), so the label is a generic top-of-self-serve marker rather than a $200 signal. Its original logic still holds where it is a single-user plan: monetise power users on priority or uncapped agentic access instead of raising the $20 mass-market price. But two boundary conditions now qualify it. First, $200 has escaped the individual at four vendors: Flora charges $200 per SEAT, Factory's Max is a seat, Adobe's Firefly for Teams Premium is $199.99 per LICENSE (2026-08-15), and Lindy converted its flat $199.99 Max into $199.99 per user on 2026-08-11 — the first observed flat-to-per-seat conversion of an existing $200 tier. Vellum's Ultra is a $200 infrastructure package (Large machine, 60 GiB, $115 credits) on an LLM-ops platform — so the number is better read as 'top of self-serve' than 'prosumer ceiling', and a buyer must check whether it is flat or per-seat before comparing. Second, the ladder gained a middle rung at ~$100 (Google AI Ultra 5× at $99.99, GitHub Copilot Max at $100/user, Factory Plus, Vellum Super, Lindy Pro, Anthropic Max from $100), which is where the volume actually sits: GitHub, with the largest installed base of any coding assistant in the corpus, caps its individual ladder at $100 and has never shipped a $200 tier. The counterexamples have also shifted from geography to strategy — Mistral ($24.99) and Suno ($30) still refuse the ceiling, but the more informative holdout is now a US frontier vendor choosing $100 on purpose.

  OpenAI Perplexity AI You.com Anthropic Cursor (Anysphere) Google Windsurf Factory Manus Hume AI Lovable Cognition Codeium Vellum Flora Lindy Adobe

  Reviewed August 2026 See the pattern

• The per-token list price is becoming a base rate times named modifiers

  Sharpens

  17 companies

  The published per-token or per-second price is increasingly the BASE of a multiplication rather than the price. Inference vendors first standardised the discount side — batch at ~0.5x (13+ vendors landing on the same number) and reduced cached-input rates — and have now built a matching surcharge ladder on the same rate card, converging on roughly 1.5x for priority scheduling, 2x for fast/high-throughput serving, and 3x for guaranteed non-preemptible execution. 12 corpus vendors now price latency or priority as an explicit axis: deepinfra (Flex 0.8x / Standard 1x / Priority 1.5x), modal (region 1.5-1.75x, non-preemptible 3x), fireworks-ai (Standard / Priority ~1.25-1.5x / Fast ~2x, plus a 10% US-only routing premium), google-gemini (Priority 1.8x, Flex-Batch 0.5x), anthropic (Fast mode 2x on Opus 4.8), together-ai (Provisioned Throughput at $0.05/PTU-minute), baseten (Mission Critical 99.95% SLA, Enterprise-quoted), voyage-ai (33% batch discount plus rate-limit tiers unlocked by cumulative spend — $100 for 2x, $1,000 for 3x), luma-ai (an unmetered relaxed/slower mode on the top tier), replit-ai (effort-based checkpoints priced on time and compute actually consumed), you-com (research_effort tiers from $12 to $1,200 per 1k calls — a 100x spread on ONE endpoint) and gladia (99.9% uptime SLA + priority queue from Growth up, Starter none). Midjourney's fast-vs-relax GPU-hour bundles are the oldest instance of the same idea. The boundary is real: sarvam-ai, sambanova and zhipu-ai publish cached-input discounts but no priority or fast tier at all, and raw GPU rental (lambda-labs) has neither. The buyer consequence is that a spec-sheet rate comparison is no longer a price comparison — a headline rate read at DeepInfra's 0.8x Flex against production traffic running at 1.5x Priority is a 1.9x error invisible in any rate table.

  OpenAI Google Fireworks AI Groq Anthropic Baseten Mistral AI OpenRouter DeepInfra Modal Together AI Voyage AI You.com Replit AI Gladia Luma AI DeepSeek

  Reviewed August 2026 See the pattern

• From subscriptions to usage — and back to hybrid, sometimes within six weeks

  Sharpens

  16 companies

  A wave of vendors abandoned flat subscriptions for pay-as-you-go metering — but several then re-introduced tier structure on top, converging on hybrid rather than pure usage. The endpoint of the migration is a platform fee or seat sitting BESIDE a meter, not pure PAYG. The strongest form of the claim is not that any single taxonomy tag leads (at 353 companies freemium 169 / subscription 160 / hybrid 155 are effectively tied) but that the pure archetypes have nearly vanished: only 10 of 353 companies are hybrid-only, 20 subscription-only and 13 pure-usage-only, while 244 (69%) carry at least one usage-bearing structure and 57 carry both 'hybrid' and 'subscription' simultaneously. The drift is bidirectional and can reverse fast — Resemble AI collapsed a five-tier ladder into one PAYG Flex plan and reintroduced Team ($280/mo) and Business ($800/mo) subscription tiers above it SIX WEEKS later (2026-07-29); Jasper reintroduced credit metering three years after killing the word caps it launched with (2026-07-22); Vellum went from a bare configurator to three named packages in under three months (2026-07-29). And a real counter-current exists: Sequence shipped an autonomous AI agent layer with no meter at all, absorbed into its existing platform fee (2026-07-21), and Anthropic bundled three more products into Claude Pro/Max/Team at unchanged prices on the same day it added $0.08/session-hour agent billing.

  Exa Tavus ElevenLabs Deepgram Lindy Synthflow Dust Qodo Jasper Linear Vellum Glean Composio Chargebee Framer ZenRows

  Reviewed August 2026 See the pattern

• Unit proliferation at the infra layer — with the corpus vocabulary now saturating and consolidation a live countermove

  Sharpens

  16 companies

  Two things previously fused must be measured separately. (1) CORPUS VOCABULARY HAS STOPPED GROWING, not merely slowed: the number of distinct billing units is 65 at 380 companies — the SAME 65 as at 353, so 27 added companies contributed ZERO new units, a marginal rate of 0.00 against 0.13 at the previous step and ~0.47 in the early passes. And the tail is now contracting rather than accumulating: single-company units fell from 15 to 13, because two former one-offs found adopters (`prompts-tracked` now at 4 companies, `sites` at 8) and no new one-off was created. A vocabulary that adds nothing and consolidates its tail is a settled language, not a saturating one. (3-bis) THE NEW COMPLEXITY IS AXES, NOT UNITS — which is precisely how a frozen vocabulary coexists with rate cards that keep getting harder. Four axes appeared in the 2026-07-31 to 2026-08-16 window and none of them created a billing unit: Fireworks' flat 1.5x for US- or Europe-pinned dedicated deployments, DeepSeek's peak/off-peak split, Sarvam's `editor_flow` request flag that exactly doubles the Dubbing rate (itself already per-minute times target-language count), and Glean's split of GPT Realtime into three modality-specific lines per model. Counting units now understates rate-card complexity. (2) PER-VENDOR METER COUNT IS STILL CLIMBING at the infrastructure and platform layer, which is the original claim and still holds: Modal, Vercel, RunPod, Glean, Composio, Fireworks, Perplexity, Anthropic and MiniMax all added metered dimensions in the 2026-07 window, and RunPod remains the most unit-diverse single vendor in the corpus (per-hour Pods, per-second Serverless, multi-node Clusters, and Public Endpoints metering by request, token, character OR megapixel depending on model). (3) CONSOLIDATION IS NOW A LIVE COUNTERMOVE, not merely an app-layer trait: roughly 4 consolidations against 7 proliferations in the 2026-07 fortnight, most starkly LangChain collapsing seven LangSmith meters into two normalized units (LCU $1.50 / LSU $1.00) on 2026-07-21 while direct competitor Braintrust went from three meters to four one day later by pricing data retention at $0.50/GB/mo. It gained its second major infra case at 380: ZenRows (2026-08-04) replaced THREE dollar-denominated bases — cost-per-1,000-requests on the Scraper API, per-GB on Residential Proxies and $0.09 per session-hour on the Scraping Browser — with one shared credits currency across four primitives at fixed published weights (1x standard, 5x JavaScript rendering, 10x Premium Proxies, 25x both), demoting Residential Proxies to '(Legacy)'. Diffbot went further and declined to proliferate at all, adding a fifth product (Web Search at 1 credit/query) onto its existing shared credit pool with no new unit and no price change. So meter count is not a one-way ratchet even within a single segment. The buyer-relevant corollary is uncomfortable: consolidation does not mean simplification. LangSmith's collapse from seven meters to two doubled the base trace rate (0.005 LSU vs the old $2.50/1k) and cut extended retention from 400 to 180 days — a shorter rate card that is harder to compare, not easier. The app-layer counterexamples still hold to one unit (Character.ai active users, Suno credits, Harvey and Superhuman seats, Fathom seats).

  Vercel Modal ElevenLabs Deepgram RunPod LangChain (LangSmith) LiveKit Weaviate Anthropic AssemblyAI Braintrust Bland AI Glean Composio Fireworks AI Perplexity AI

  Reviewed August 2026 See the pattern

• Deal-desk and billing-engineering hiring marks monetization maturity

  Holds

  15 companies

  Whether a company is hiring deal-desk, RevOps, and billing-engineering roles is a leading indicator of where it sits on the monetization-maturity curve. The signal escalates, and it has now been tested FORWARD: both predictions logged at the 2026-07-15 review landed within 16 days — HubSpot's June-2026 Principal Engineer - Commerce req shipped as the priced Revenue Hub on 2026-07-30 (roughly six weeks from req to monetized product), and Freshworks' VP Billing & Renewals plus Director Pricing & Packaging pairing produced a published Freddy AI Agent rate ($49 per 100 bot sessions) on 2026-07-21 after two years of sales-quoted session packs. Recounted at 380 companies, where the hiring base did NOT grow — 176 companies carry hiring data at 380, the same 176 as at 353, so coverage fell from 50% to 46% while every count stayed identical (263 companies still carry a classified stack entry, 69% down from 74%). Companies-with-at-least-one-req run: RevOps 127 (near-universal once a revenue org exists), customer-success ~91 (a broad post-sales bench; published as 118 and corrected on 2026-08-17), growth 78, a dedicated pricing/monetization role 63, billing-eng 62, deal-desk 49, retention ~25 (published as 69 and corrected on 2026-08-18 — the same boilerplate defect, compounded by 'retention' and 'renewal' usually denoting data retention, employee retention or a benefits renewal), data-platform 21, and cost-finops ~10 (published as 12). Those counts were carried forward, not re-confirmed, at the 2026-08-16 review: nothing in the 2026-07-31 to 2026-08-16 window added a single company to the hiring base, so that cycle tested the trend's THREAT rather than its claim. The re-confirmation ran on 2026-08-17/18, and it was a MEASUREMENT re-test rather than a fresh board sweep: three rungs turned out to have been counted by matching keywords anywhere in a job description, which is why customer-success, retention and cost-finops are restated above. Two things changed at the earlier 353 recount from the stale 190-block base. First, the ladder's order inverted: a pricing/monetization role (63) now edges out billing-engineering (62), reversing the prior 59-vs-51 ordering — at 353 companies the packaging question (seats vs usage vs outcomes) arrives at least as early as the plumbing. Second, cost-finops enters the ladder at ~10 companies, a margin-pressure cohort to watch as inference becomes the dominant COGS line. Deal-desk remains the most diagnostic signal at 49 of 176 (28% of companies with hiring data) — it is no longer the rarest rung, since corrected retention (~25) now sits below it — it surfaces only when a company negotiates non-list-price enterprise deals at enough volume to need a quote-approval function — and companies that open one are almost always layering a sales-led enterprise motion onto a self-serve base, which is why they also carry the heaviest customer-success benches. (The matching claim about RETENTION benches did not survive the 2026-08-18 correction: a dedicated retention function is rare corpus-wide.) (All counts are companies-with-the-role, not headcount.) The forward risk to the signal is commoditization: four corpus vendors productized CPQ inside twelve days (HubSpot's Revenue Hub Professional from $95/mo, Chargebee's CPQ Lite free for 50 quotes, Freshworks' CPQ at $19/user/mo, Sequence's quote-builder add-on), so if quote-approval capability becomes a line item, the deal-desk req may stop being the tell and the pricing/monetization req may have to carry the signal. That threat escalated at 380: Sequence (2026-07-21) shipped a QUOTE-APPROVAL AGENT — alongside AR, reconciliation, invoice-review, payment-reminder and contract-intake agents, reviewed in a Watchtower console — with no new priced SKU and Growth still at $799/month. A free agent removes the headcount rather than cheapening the tool, which erodes the deal-desk tell faster than a $19 line item does. The counter-signal remains the lean self-serve company that meters in-house but opens no RevOps, deal-desk or billing-eng req at all.

  OpenAI Anthropic Clay Decagon Stripe Billing Baseten Harvey Cohere Coreweave Sequence Nebius Modal Socket Waymo Parloa

  Reviewed August 2026 See the pattern

• The meter tracks the buyer: usage for developers, commitments for enterprise

  Holds

  15 companies

  Two variables, nested. The strong one is the UNIT: a unit a machine emits can be metered per event and therefore can be sold pure-usage, while a unit a human occupies collapses to a subscription. Six machine-metered units land 70–88% pure-usage against a 32% corpus base (gpu-hours 88%, requests 76%, characters 75%, api-calls 72%, tokens 71%, cpu-hours 70%); seats land 13%; active-users land 71% seat-based against a 21% base; credits land in neither pole, attaching to freemium (67%) and hybrid (59%) because a credit is an abstraction layer over a unit rather than a unit. You can therefore predict a vendor's pricing model — and your budget's variance — from the unit on its rate card alone. The weaker, older variable is the BUYER: commitments remain an enterprise artifact (154/353 have commits, 44%; infra-cloud runs 28/42 at 67%), and the same vendor routinely sells the same compute two ways depending on who signs (RunPod's self-serve PAYG vs its 2026-07-22 Enterprise reserved-baseline-plus-burst; Together's on-demand H100 at $5.49/hr vs reserved at $4.99). But that half now has a live exception: Hyperbolic made reserved capacity self-serve and prepaid in 1-week terms on 2026-07-21, so commitment is starting to detach from enterprise procurement. The most informative cell is the 21 companies that bill in tokens yet are NOT pure-usage — 8 of them (augment-code, codeium, continue, cursor, tabnine, v0, windsurf, bolt-new) buy tokens and resell seats, absorbing the usage risk their customers refuse to carry.

  DeepSeek Modal Fireworks AI Together AI Anyscale Glean Harvey Vercel Lambda Labs Vapi LiveKit RunPod GitLab CoreWeave Nomic

  Reviewed August 2026 See the pattern

• AI usage graduates onto its own dedicated meter, decoupled from the base plan

  Sharpens

  14 companies

  Vendors are pulling AI usage onto its own dedicated billing meter — a separate credit currency, marketplace consumption unit, per-token platform fee, or agent multiplier — priced independently of the base product's seat, plan or edition. The driver is unit-economics honesty: as AI inference becomes a large, volatile cost-of-goods line, vendors want its price to float on its own axis (routing, model, marketplace) rather than inherit the base plan's tiering, so a cheaper model or a cross-region route reaches the meter without re-rating the whole account. The pattern started at the platform layer (Snowflake's edition-independent AI Credit, Vercel Agent's $0.25/1M, Anthropic's $0.01 Claude Consumption Unit, W&B's ARIA line, Google's 2x AlphaEvolve surcharge) and this trend set itself an explicit test: it would remain a data-warehouse/dev-platform quirk unless app-layer, seat-based vendors followed. As of 2026-07-30 they have. Linear moved a second AI feature ('Loops') behind purchased AI credits at unchanged $10/$16 seats (2026-07-29); Jasper added a credit meter and a public rate card to a previously meter-free per-seat product (2026-07-22); Freshworks published a per-unit AI rate on a seat-based CRM ($49/100 bot sessions) with seats flat (2026-07-21). Modal made the platform-layer version of the same move at the unit level, launching a token-metered Shared API alongside — not instead of — its per-second GPU rate card (2026-07-29), and meters keep multiplying (Braintrust's retention became a fourth meter at $0.50/GB/mo; Composio's August model breaks LLM tokens out from tool calls). The countervailing bundling force has NOT weakened and is winning at the top: Anthropic folded three new products — Claude Cowork, Design and Science — into Pro, Max and Team at unchanged prices (2026-07-23), Sequence shipped five named revenue agents with no priced SKU at all, and GitHub Copilot and Shortwave still sell AI capacity only by moving up a tier. The live question is no longer whether the pattern reaches the app layer but which force wins per vendor — and Linear's own pricing page, with two AI features metered and two bundled, shows a single vendor can run both at once.

  Snowflake Cortex Vercel Anthropic Frase Weights & Biases Google Linear Jasper Freshworks Modal Braintrust Glean Framer Mintlify

  Reviewed August 2026 See the pattern

• Billing-infra vendors gate their own pricing

  Holds

  14 companies

  The gating clustering in billing infrastructure is real but shrinking, and the ironic framing that named this trend no longer fits. 9 of 17 core billing/payments vendors in the corpus are gated or sales-only (53%) — still 1.8× the 29% corpus baseline, but down from a 100% rate at first observation and from ~60%+ at the last review. Two of the segment's most prominent vendors crossed over: Metronome now publishes Starter rates (0.8% of billing volume, $0.04/1,000 events after $100K and 10M events included) after Stripe closed its acquisition on 2026-01-14, and Chargebee published a four-tab public rate card on 2026-07-22. The remaining gated set (Kill Bill, Lago, m3ter, Maxio, OpenMeter, Orb, Sequence, Togai, Zenskar) is dominated by pure metering engines whose own price is a function of the customer's billing volume and event count — the structural reason a simple rate card is hard, and the reason m3ter's page still shows a four-step quote with zero dollar amounts even after Salesforce took ownership on 2026-07-21. The sharper claim to carry forward is not that billing vendors hide prices, but that the segment splits by GO-TO-MARKET rather than by product: PLG-first and open-source entrants publish (Flexprice, Schematic, Alguna, Hyperline's Launch tier, Metronome's Starter, Stripe Billing at 0.7%), while enterprise-quote entrants do not. And the buyer-side context matters more than either: across the 263 corpus companies with a documented stack, billing and payments are bought 9.4:1 over building, while metering is built 4.8:1 over buying — the market these vendors sell into mostly writes its own meter.

  Lago Metronome OpenMeter Orb Togai m3ter Chargebee Maxio Zenskar Sequence Finout Kill Bill Hyperline Flexprice

  Reviewed August 2026 See the pattern

• Dollar floors go dark, burn rates go public

  New

  14 companies

  Disclosure is not retreating or advancing; it is re-sorting by which half of the price equation gets published. Ten corpus companies removed a previously published price and fourteen published one for the first time in the same window. The withdrawn items are overwhelmingly dollar-denominated floors and ceilings — Netlify's $500/mo Enterprise floor, Moonshot's $599/yr/seat, Arize's ~$10/M-span overage, Metronome's $100k included allotment, Usage AI's take-rate percentage. The newly published items are overwhelmingly per-ACTION burn rates — Apollo's per-endpoint credits, Glean's 13-capability FlexCredit card, Jasper's per-action credits, Luma's and Photoroom's absolute allowances, Lindy's credit bands. Vendors are trading knowability of the BILL for knowability of the METER, and the combination is strictly less useful to a buyer than either half alone, because it lets a vendor reprice what a credit costs without touching a number anyone can diff. Boundary: three disclosures break it by publishing plain dollars (Zilliz, Freshworks, You.com). Falsified if a vendor publishes an enterprise floor while withdrawing a burn rate.

  Apollo Glean Jasper HeyGen Luma AI Lindy Netlify Moonshot AI Metronome Arize AI Groq Usage AI Poolside Observe.AI

  Reviewed August 2026 See the pattern

• Expiring generosity: unlimited plans and promo rates with printed end dates

  Sharpens

  14 companies

  Time-boxed generosity has moved from promo copy to the primary rate card, and now spans 11 of 353 corpus companies. Two distinct forms exist and should not be conflated. (1) Dated discounts — Intercom's 35%-off Essential, Tavus's 'first month 50% off', Oxylabs's WU40, Puzzle's '100% off for 2 months', HubSpot's 60-day fee waiver — are acquisition levers; they revert quietly and buyers should model list, not the badge. (2) Dated entitlements and dated list prices — Genspark's unlimited-through-2026-12-31, Anthropic's Sonnet 5 at $2/$10 'through August 31, 2026' then $3/$15, Braintrust's Pro credit printed to fall from $249 to $100 on 2026-09-01, W&B's ARIA free 'for a limited time' — are pre-authorized price increases on the core value promise, and are the consequential half: they let a vendor hedge model-cost drift without ever announcing a repricing. The first four resolutions in the corpus show all three possible endings actually occur: quiet renewal (Braintrust deleted its own July 31 clause one day early), on-schedule reversion (Intercom, Tavus), and conversion into a permanently higher list price (Zhipu, +26% to +40%). So the printed date is real optionality, not boilerplate — and because it is optionality, a buyer cannot infer the outcome from the clause. The live test is Anthropic's 2026-08-31 Sonnet 5 sunset: a frontier lab honouring a printed 50% increase on flagship tokens would make the mechanic standard practice.

  Genspark Deepgram Oxylabs Intercom Anthropic Weights & Biases Tavus Puzzle Zhipu AI Braintrust HubSpot DeepSeek Fireworks AI FLORA

  Reviewed August 2026 See the pattern

• Retrieval and data-API vendors widen free tiers to grab evaluation

  Sharpens

  14 companies

  Retrieval, scraping, extraction and vector-DB API vendors are competing on the width of the free/evaluation entry, enlarging recurring free tiers and cutting entry floors to win developer land while planning to monetize at production volume. The pattern held through a second cycle: after the 2026-07-14 batch (Bright Data's recurring no-card tier, Exa's 20x, Powerdrill's daily pool, Unstructured's monthly reset, turbopuffer's 75% floor cut), seven more retrieval/dev-API vendors widened in the following fortnight — You.com added 100 free calls/day to its flagship Web Search API, Firecrawl bundled a flat +1,000 search credits into every plan including Free, Exa widened again and re-denominated its free tier into dollars, Socket 10x'd its Free dependency cap, Apify doubled Free RAM, LlamaIndex lifted Free seats to 100, and Krisp opened its first self-serve API door. The boundary is now sharp and it validates the stated mechanism: where an evaluation request is near-free to serve, the free tier widens; where it burns GPU or audio time, it retracts. In the same window Cerebras, Gladia and Novita all converted recurring free access into one-time credit grants, and six app-layer vendors (Gumloop, Freshworks, Maxio, Photoroom, Puzzle, Playground) cut or killed free tiers outright — with corpus-wide free-tier prevalence at 215/353 (61%), down from 64% at the 338 base. The barbell against premium gating is real but not universal: Bright Data raised Bright Insights 5-8x and Linkup's Extract needs a $10 balance, yet You.com cut its top Research tier from '>$2,000' to a firm $1,200 per 1k calls a week before adding its free tier. Open question: whether retrieval free tiers survive their own success once evaluation traffic becomes production traffic.

  Bright Data Exa AI Powerdrill Unstructured turbopuffer You.com Firecrawl Socket Apify LlamaIndex Krisp ZenRows Diffbot Frase

  Reviewed August 2026 See the pattern

• AI-coding tools pair a seat with a meter — but the meter itself has a 12-month half-life

  Sharpens

  13 companies

  AI-coding tools pair a per-seat subscription with a usage meter, but the SHAPE of that meter is the least stable thing in the corpus. Recounted at 353 companies: 30 in-corpus companies carry use_case ai-coding or code-generation, and 22 of the 30 pair seats with at least one usage unit — yet only 12 of 30 carry CREDITS as a billing unit (cognition, continue, cursor, dify, factory, github-copilot, gitlab, replit-ai, sourcegraph-cody, sweep-ai, v0, windsurf), 8 meter raw tokens next to the seat (anthropic, augment-code, bolt-new, claude-code, codeium, openai, poolside, tabnine), 2 keep credits with NO seat (lovable, qodo), and 2 run flat quotas with no live meter at all (imbue, phind) — plus factory, which is the one genuine double-count: it still carries `credits` in its billing_units tag but its live mechanic as of the 2026-07-22 capture is per-tier rolling rate limits, not a credit meter, so it appears in the credits row on its taxonomy and in the quota camp in practice. aider charges nothing. The segment is the corpus's densest concentration of the seat-plus-usage taxonomy — 10 of the 32 seat-plus-usage companies corpus-wide are AI-coding tools. What the last 14 months actually show is a 4-4 standoff on the meter: four vendors moved OFF credits onto quotas or rate limits (Factory 2026-04 after a year of token metering; Windsurf 2026-04, its third core metering change in ~16 months; Claude Code's weekly limits 2025-07-28 then doubled 5-hour limits 2026-05-06; and in the adjacent agent-builder segment Lindy 2026-03-10) while four moved ONTO credits (Cursor 2025-06-16, followed by a public apology and refunds 18 days later; Augment Code 2025-10-20; Cognition's ACU pay-as-you-go 2025-04-03; Qodo 2026-06-30). GitHub Copilot is the instructive third path — it chose credits AND kept the seat (2026-06-01), and its credit allowance now exceeds its seat price at every tier ($10 seat / $15 credits, $39 / $70, $100 / $200), so the seat is no longer the revenue anchor. A third metering axis appeared on 2026-07-23: Claude Managed Agents bills $0.08 per session-hour alongside tokens. The practical consequence: in this segment the meter has a demonstrated ~12-month half-life, so an internal chargeback model built on any single coding agent's current unit will be wrong within a year.

  Cursor (Anysphere) Codeium GitHub Copilot Augment Code Qodo Windsurf Sourcegraph Cody Tabnine Bito Continue Cognition Claude Code v0

  Reviewed August 2026 See the pattern

• BYO-API keys zero out the model-cost meter

  Sharpens

  13 companies

  Agentic and workflow platforms split the bill into two meters — a platform/orchestration credit they keep, and a passthrough model/data cost they mark up — then offer Bring-Your-Own-Key as a lever that zeroes or halves the passthrough meter. BYOK makes the vendor's role explicit: you pay them for orchestration and pay the model provider directly for inference, which is a transparency and margin signal (the platform concedes it is not the cost centre on inference). The lever still works wherever the split is published — Clay's BYOK eliminates Data Credits but still burns Actions, Relevance's zeroes Vendor Credits, Gumloop's halves AI-model credits, Tabnine's makes token usage unlimited against a published 5% handling fee, and Bland AI added a second BYO-carrier case on 2026-07-21 by moving telephony out of its all-in per-minute rate at unchanged plan prices. But the boundary condition identified on 2026-07-30 is that the SPLIT ITSELF is now contested, and that is a bigger threat than vendors withdrawing BYOK (Cursor's route). Three ways the precondition erodes: a vendor can MERGE the currencies so the passthrough meter is no longer visible (Lovable folded its dollar-denominated AI-gateway and Cloud balances into one credit balance), it can concede WHOLESALE so the lever buys nothing (Perplexity resells 7 providers' models 'at direct provider rates with no markup'), or it can give away BUILT-IN model access so no key is needed (Braintrust deleted GLM-5.2's sunset date one day before its own cutoff). BYOK is therefore best read as a diagnostic for whether a passthrough markup exists at all — where the meter is merged or the markup is zero, the lever is not withdrawn so much as beside the point.

  Clay Relevance AI Gumloop Byword Vectara Lemlist Tabnine Deepgram Synthflow Aider Continue OpenRouter Bland AI

  Reviewed August 2026 See the pattern

• Self-serve checkout is the transparency forcing function

  Holds

  13 companies

  Price transparency in this corpus is set by whether a self-serve purchase path exists, not by whether a sales team exists — 230 of 335 sales-led companies publish prices (69%, against a 72% corpus base), while removing the self-serve/PLG path drops publication to 3 of 78 (4%). Both rates are unchanged from the 353-company review (68% vs 71%; 3 of 73), so the claim survived a 27-company expansion without moving. The enabler is mechanical: a checkout cannot take money without a number on the page, which puts this closer to a definition than to a discovery. What keeps it empirical is that the SAME 3 companies — Lorikeet, Regie.ai, Tabnine — publish a rate card they refuse to sell self-serve at both 353 and 380; the corpus grew and the exception set did not. Two boundary conditions. First, price_transparency is a single-valued field on multi-product companies, so the 4 apparent reverse cases (automation-anywhere, clockwise, sana-ai, unbabel) are product-line splits rather than counterevidence. Second, a wider cut counts 7 companies publishing without a self-serve tag (adding docket, shopify, vectara, yellow-ai), but all four sit in a `plg`-without-`self-serve` motion cell and at least two are mis-coded — Shopify publishes Basic $39 / Grow $105 / Advanced $399 / Plus $2,300+ and plainly transacts self-serve — so that number should be audited, not cited. CORRECTION: the 2026-07-30 claim that sales-led + self-serve without PLG publishes BELOW base (33/50, 66%) no longer holds. At 380 the cell is 41 of 58 (71%), level with the 72% base; all eight companies added to it publish. A self-serve path without PLG now predicts publication as well as the corpus average does.

  Harvey Glean Gong Ada Sierra Mercor Skydio Orb m3ter ZoomInfo Lorikeet Regie.ai Tabnine

  Reviewed August 2026 See the pattern

• AI ships bundled, infrastructure gets unbundled

  New

  12 companies

  Two opposite packaging moves ran at equal volume in one window, and the split is by what the thing costs the vendor to run. Nine companies shipped new AI capability into existing tiers with no price change — Anthropic bundling four sibling products into unchanged Pro/Max/Team subscriptions, Wispr Flow shipping an entire second product free, FLORA putting a new studio on every plan including Free. Nine others carved out something previously bundled and priced it: Bland AI removed telephony from its all-in per-minute rate, Braintrust turned data retention from a feature gate into a $0.50/GB/mo meter, Stripe moved advanced usage billing out of its 0.7% rate into Metronome, CrewAI removed 50 monthly development hours from Enterprise contracts. The vendors are giving away the input with the highest apparent marginal cost — GPU inference — and charging for storage, telephony minutes, trace egress and human hours. The boundary is a tendency, not a rule: Surfer SEO unbundled AI-visibility tracking into a standalone slider-metered product, and Vercel priced net-new AI telemetry rather than bundling it. Falsified if standalone pricing on net-new agent capability becomes the majority behaviour.

  Anthropic Wispr Flow FLORA Sequence Diffbot Mem0 Firecrawl Bland AI Stripe Braintrust CrewAI Zendesk AI

  Reviewed August 2026 See the pattern

• Clinical AI converges on per-clinician gated contracts

  Holds

  12 companies

  Clinical AI pricing is not uniform — it stratifies by who funds the purchase, and that variable predicts both the billing unit and the transparency level. (1) Health-system operating budget → per-clinician (PEPM) annual contract, sales-gated: Abridge, Ambience, Suki, Rad AI, Viz.ai and Nabla Enterprise, driven by EHR integration economics (Epic and Cerner charge by enrolled clinician), FDA/HIPAA compliance attaching to the provider seat, and multi-stakeholder IT procurement. (2) Individual clinician's own card → per-seat, public, self-serve: Freed, Heidi Health, Nabla's free tier, Eko+ at $12.99/mo. (3) Payer/CMS reimbursement → per-procedure with a PUBLISHED price, because the government publishes it: Eko SENSORA at CPT 0962T (~$128.90), Viz.ai LVO NTAP (~$1,040/use), Paige ~£1/slide via NICE, Tempus xT CDx ADLT $4,500/test. So 6 of 11 clinical AI companies are sales-only and 8 of 11 are non-public — high, but not the 100% originally claimed. Two newer forces are pulling the PEPM unit apart from above: as the AI does more of the documentation work seat count stops tracking value, so Abridge (2026-07-22) now names an unlimited system-wide agreement that decouples price from headcount, and Hippocratic bills per agent-hour (~$9/hr vs ~$39/hr RN wage) rather than per seat. Expect the health-system branch to migrate from PEPM toward site licenses and per-encounter/per-outcome units, while the reimbursement branch stays the most transparently priced clinical AI in the corpus.

  Abridge Eko Health Tempus Ambience Healthcare Suki AI Nabla Freed Heidi Health Paige AI Rad AI Viz.ai Hippocratic AI

  Reviewed August 2026 See the pattern

• Daily-refresh credit pools on prosumer AI plans

  Holds

  12 companies

  Prosumer and agentic AI apps layer a small daily-refresh credit drip onto (or beneath) the monthly pool — 300/day at Manus, ~10/day at Higgsfield free, 100/day at Genspark free, $2/day on v0 paid seats, daily refresh on Julius's annual pools — because the cadence does two things at once: it caps the worst-case daily blast radius of an expensive agent run, and it nudges habitual engagement (use-it-or-lose-it). As of 2026-07-30 the cadence is no longer a daily-vs-monthly binary but a five-point ladder: sub-daily rolling windows (Playground's 5 images per 3 hours), daily (v0, Lovable, Manus, Genspark, Higgsfield, Suno, Powerdrill, Poe's $4.99 tier), daily-AND-weekly (Cognition's Devin), monthly (most), and annual-plus-daily (Julius). Two boundary conditions are now documented. The mechanic has broken out of the free tier — Powerdrill runs it on every paid tier, Poe runs it on its cheapest PAID tier while monthly pools sell the expensive ones, and Cognition's Devin is the first corpus case where a daily+weekly allowance REPLACED the credit pool outright rather than layering on it. But it is not converging into a standard: Kaiber made credits roll over indefinitely, Hyperbolic's credits never expire, and n8n delivered a 46-91x allowance increase on a monthly cadence with no rollover — so where vendors got more generous in mid-2026, they mostly did it monthly.

  Manus Julius AI Genspark Higgsfield v0 Lovable Suno Powerdrill Cognition Poe Playground AI Adobe

  Reviewed August 2026 See the pattern

• Per-action pricing for agent tools

  Sharpens

  12 companies

  As agents call tools — web search, page visits, code execution, sandboxes, deep research — vendors price those actions individually, per call or per hour, alongside and on top of tokens. The unit is established: at least 7 inference and search vendors now publish a server-side tool rate card (Groq, xAI, Perplexity, MiniMax, Browserbase, Upstash, Modal), and on agentic workloads the tool line can dominate the token line — xAI meters Web Search, X Search and Code Execution at $5/1,000 calls each, sitting above its own per-million-token rates in cost terms for a chatty agent. As of 2026-07-30 the sharper claim is that the action unit has become the volatile part of the bill rather than a stable add-on. It now moves in every direction, fast: Perplexity halved web_search ($0.005 → $0.0025) and fetch_url ($0.0005 → $0.00025) eight days after adding a $0.03-per-session sandbox meter; Composio is taking tool-call overage from $0.299/1K to $4/1K (13x) on 2026-08-15 while cutting Pro's included calls 200K → 50K; and Groq withdrew a priced Browser Automation tool ($0.08/hour) seven days after launching it, the shortest priced-SKU life in the corpus. The unit is also fragmenting in shape — flat per-call, per-hour, per-session-with-a-billing-window (Perplexity, and OpenAI's Containers from 2026-03-31), per-minute (Glean's Meeting Notes), a multiple of the underlying token rate (Google's AlphaEvolve at 2x), and surface-dependent (Glean charges 1 FlexCredit for a Basic Search Query via the Client API and 0 for the identical query in-product). Counterexamples are narrowing: OpenAI published Containers tool pricing on 2026-07-23, leaving Mistral, Cohere and DeepSeek as the vendors shipping agentic models with no separate tool SKU.

  Groq Perplexity AI You.com Exa Anthropic Browserbase Composio Zapier Vapi MiniMax Zhipu AI Glean

  Reviewed August 2026 See the pattern

• The pricing page is an experiment surface

  New

  12 companies

  A meaningful slice of the corpus now treats the pricing page as a surface to test on rather than a statement to stand behind. Twelve companies reversed a pricing or packaging decision within 30 days in a single 31-day window. The reversals concentrate in promos, tier LAYOUT and free-grant structure rather than in per-unit rate cards: Oxylabs flipped its page layout three times while every per-1K rate stayed byte-for-byte identical, and Anthropic held $5/$25 across five model generations. The consequence is measurement, not morality — "current price" becomes a probabilistic statement, and a competitive rate table built from one weekly snapshot has a material chance of citing a value that was already gone. It also contaminates any metric that counts packaging events, because a round trip inflates the count without changing the menu. Falsified if reversals start landing on published per-unit rates at the same cadence, or if the rate falls below roughly 3 per month in a later window.

  Novita AI FLORA Hyperline Tavus Oxylabs Perplexity SambaNova Metronome Intercom Wispr Flow Make Resemble AI

  Reviewed August 2026 See the pattern

• Incumbent CRM/sales SaaS bolts an AI credit meter onto the seat

  Holds

  11 companies

  The AI transformation of incumbent, seat-licensed CRM and sales-engagement SaaS is arriving as a bolt-on consumption meter, not a repricing. 8 of the 13 CRM/sales incumbents in the corpus keep the per-seat license as the revenue anchor and add a separate AI credit/consumption currency on top — HubSpot Credits, Microsoft Copilot Credits, Close's Chloe credits, Freshworks' per-session Freddy meter, and the seat+credit pools of ZoomInfo, Outreach and 6sense (plus Keap's usage add-ons). The 5 holdouts (Pipedrive, Copper, Zoho, SugarCRM, Salesloft) are mid-market pure-seat CRMs that fold AI into the edition, so the split tracks category — horizontal platforms and data-heavy sales-intelligence tools meter; pipeline CRMs bundle — not company size. Two refinements as of 2026-07-30. First, the overlay is not a single extra meter but a ladder: HubSpot now runs seat → AI credit → payment volume, adding Revenue Hub with a 2.9%+$0.50 card fee (and a 0.75% tax on sellers who keep Stripe), and Salesforce bought m3ter's metering-and-rating engine outright. Second, the overlay is becoming transparent even as it deepens — Freshworks published its Freddy rate ($49/100 bot sessions) after two years of quoting it, Apollo published a per-endpoint credit table showing the pool spans human in-app reveals AND API calls (mobile numbers cost +8 credits), and HubSpot deleted the entire Breeze brand without moving one per-outcome price. The credit still does its original double duty — keeping the familiar seat contract intact while capturing volatile AI cost on a repriceable meter — but its dollar rate is increasingly published, so the opacity argument is weakening faster than the architecture is. The live alternative is Zendesk's route: raise a per-agent AI add-on 66% ($50 → $83) and sell minutes in blocks, never introducing a credit at all.

  HubSpot Microsoft Dynamics 365 Close CRM Freshworks ZoomInfo Outreach Keap 6sense Apollo.io Klaviyo Salesforce

  Reviewed August 2026 See the pattern

• PLG reprices in public, sales-led reprices in private — the visible velocity gap

  Sharpens

  11 companies

  Visible pricing velocity tracks go-to-market motion, but the size of the gap and the completeness of the invisibility both needed correcting at the 2026-07-30 review. THE GAP, measured on the first window large enough to normalize (101 customer-facing events, 2026-07-15 to 2026-07-30): 84 events from plg-tagged companies, 9 from self-serve-only, 8 from sales-led-only. Raw that is 10.5:1, not the ~20:1 previously cited. Normalized against corpus composition it is smaller still — plg-tagged companies are 63.5% of the corpus and produced 83.2% of events (1.31x over-index), sales-led-only companies are 20.7% and produced 7.9% (2.6x under-index), so event rate per company differs by roughly 3.4x. A meaningful share of the original 20:1 was corpus composition, not behaviour. THE MECHANISM, now better specified: PLG pricing is public, self-served and A/B-iterable, so it gets restructured in the open and often (VEED was caught running a live $22-vs-$24 A/B on its recommended tier across 16 loads from one egress). Sales-led pricing changes happen inside contract renewals — but they do NOT leave no trace. They surface through two channels this trend had not identified: hyperscaler MARKETPLACE LISTINGS and ENTERPRISE RATE CARDS. Observe.AI, price_transparency sales-only with a 404 on its own pricing page, published $4.80 per Minutes-per-Month unit and $12.00 per Interactions unit on AWS Marketplace (2026-07-22); Glean, gated, expanded its Enterprise Flex rate card to 13 metered capabilities (2026-07-22). Sales-led vendors also disclose STRUCTURE more readily than price — Abridge named an 'unlimited enterprise agreement' without a number, Inflection retired its Enterprise offering to a 404. Note too that GTM tag and transparency are separate axes: 3 of the 73 sales-led-only companies publish prices, and Lorikeet (sales-led motion, public transparency) shipped a real mechanic change in this window. Buyer takeaway, revised: if you buy PLG, expect your plan to be restructured under you within ~12 months. If you buy sales-led, expect a stable-looking page and per-renewal negotiation — but check the vendor's marketplace listing before concluding the price is unpublished. The rate card you can see is not the cadence you will experience, and the pricing page is no longer the only place to look.

  Corpus tally Perplexity Windsurf Dust HeyGen You.com Batch tally Window tally Mercor Glean Lorikeet

  Reviewed August 2026 See the pattern

• The seat stops being the price axis

  Holds

  11 companies

  A minority of AI-native vendors are deleting the seat outright — replacing the per-user fee with a flat per-org or per-workspace price plus a consumption meter — rather than stacking credits on top of seats. 11 of 380 corpus companies have done it, and the stated enabler is that agents, CI runners and pipelines are not headcount: the per-seat charge Netlify removed on 2026-04-14 counted active Git contributors, and its CEO framed the replacement around AI-paced development "rather than human headcount." The boundary condition is that this is a live counter-current inside a still-seat-dominated corpus, not a completed shift — and at 380 that boundary hardened on every measure. Seats remain the most common billing unit at 205 of 380 (54%, up from 53%); 17 of the 27 companies added this cycle bill seats, a 63% marginal rate; 97 of the 149 credit-billing companies (65%) also bill seats, up from 88, so the layering-over-removal ratio widened from 8-to-1 to 8.8-to-1. Resemble AI reversed its own seat-free move in seven weeks, and Lindy (2026-08-11) ran the opposite conversion outright — a vendor with NO seat axis moved every tier onto per-user billing while cutting its entry price 40%. If three or more of the 11 restore a per-seat charge within twelve months, or their flat fees start rising faster than seat prices in the same category, this is repackaging optics rather than structure. As of 2026-08-16 that count is ZERO of three, with zero new adopters in the same period: the cohort is stable, not growing.

  Netlify Helicone lemlist Qodo LlamaIndex Gumloop Lovable Braintrust Langfuse Alguna Schematic

  Reviewed August 2026 See the pattern

• Billing vendors build what they sell — but only a third can prove they bill themselves on it

  Holds

  10 companies

  Two claims were previously fused and need separating. CLAIM A (near-universal): a company whose product is a usage-billing or metering platform builds that engine itself and never buys a rival's. Of the 13 such vendors in the corpus carrying a monetization_signals block (alguna, flexprice, lago, m3ter, maxio, metronome, openmeter, orb, schematic, sequence, stripe-billing, togai, zenskar), 10 disclose an in-house metering/billing/entitlements engine as a stack entry and ZERO name a competing billing vendor as their own system of record. CLAIM B (much narrower): the vendor actually runs its own money through that engine. Explicit first-person evidence exists for only 4 — openmeter ('and yes, we use OpenMeter to bill for OpenMeter'), lago ('logically decided to use Lago to craft and power our own custom billing'), sequence (its own finance-stack req names 'Sequence (Accounts Receivable & Quote to Cash)'), and metronome, whose case strengthened materially on 2026-07-14 when it published Starter rates priced on the exact two dimensions it meters for customers (0.8% of billing volume + $0.04 per 1,000 events after $100K/10M included). Two more assert it without a first-person quote: stripe-billing (direction: 'internally it builds and dogfoods that stack — Stripe Billing, Invoicing, and Revenue Recognition', but the sourced quote is product reach) and m3ter, which prices its own gated quote on the two dimensions its engine meters ('usage data ingested and bill calculations performed'). Four disclose only the build, never the self-billing: flexprice, maxio, schematic, togai. That is the full 13 — 4 proven, 2 asserted, 4 build-only, 2 counterexamples, 1 unknown. Two vendors break the pattern outright: orb's only named monetization tool is still Salesforce even though it is the bought biller in 8 other corpus stacks, and zenskar runs no self-serve at all and quotes every deal — the legacy motion its product is sold against. A fourth structural fact now dominates the cohort: 4 of the 13 have been acquired by a larger payments/CRM platform (togai→Zuora 2024-04-30, metronome→Stripe Jan 2026, orb→Adyen $335M 2026-06-11, m3ter→Salesforce as of the 2026-07-21 capture), and in every 2026 case the acquired vendor's own pricing page survived the deal word-for-word.

  Lago Sequence Metronome OpenMeter m3ter Flexprice Maxio Schematic Togai Stripe Billing

  Reviewed August 2026 See the pattern

• Flat fee, banded by your scale

  Weakens

  10 companies

  The billing-infrastructure and FinOps cohort is replacing percentage-of-revenue and percentage-of-spend take-rates with flat fees gated by a band of the customer's own scale — revenue, billings, tracked cloud spend, active contracts, billable end-customers, even website visitors — and four vendors now market "not a percentage of your revenue" as the headline claim. Hyperline's 2026-07-29 repricing is the first dated exit from a take-rate in the corpus, and only 2 of the 17 billing-tagged companies still publish a percentage (Stripe Billing 0.7%, Chargebee 0.75%). The mechanic is not free of percentages, though: Sequence's quoted tiers are still "a monthly platform fee on projected billed revenue," Finout's flat bands are reported to land at roughly 1% of the cloud bill, and Vantage keeps a savings-share on its Autopilot product while refusing one on its subscription — so the percentage often moves into the quote or onto the outcome rather than disappearing. If any of the ten restores a published take-rate — Orb already did the reverse move in 2024-11, adding billings as a primary metric after advertising that it charged no percentage of billings — or a new entrant wins share at a sub-0.5% rate the flat-fee cohort has to match, the "percentage is dead in billing software" claim breaks.

  Hyperline Kill Bill Zenskar Vantage Sequence Maxio Flexprice Finout Schematic Docket

  Reviewed August 2026 See the pattern

• The agent sandbox meter has not converged

  Holds

  10 companies

  The agent's execution sandbox has become a billable unit at 9 of 353 corpus vendors, but nobody agrees what the unit is: 6 distinct shapes are live simultaneously and only per-vCPU-time plus per-memory-time has more than one adopter. The enabler is that agents now hold long-lived stateful compute, which neither a token meter nor a request meter prices. The divergence is structured, not random — infrastructure-shaped meters bill for capacity held while session-shaped meters bill for agent time — and the two are not inter-convertible without knowing a sandbox's core count and duration, which is exactly what a buyer does not know before deployment. The open question is which family wins: if the session shape spreads, the sandbox becomes its own unit; if vendors fall back to vCPU-time, it was never a new unit at all, only cloud compute at a multiplier, which is what Modal already says out loud at roughly 3x its own standard function rates.

  Perplexity Anthropic Modal Together AI Novita AI Upstash LangChain / LangSmith OpenAI Vercel Composio

  Reviewed August 2026 See the pattern

• The credit equation is published in halves

  Weakens

  10 companies

  Vendors that price in a synthetic unit — credit, FlexCredit, LCU, ACU, DBU, compute point — systematically publish only one half of the two-part equation a buyer needs to reach a cost per action, either dollars-per-unit or units-per-action but almost never both, which makes the unit an opacity device rather than the simplification it is sold as. Ten corpus companies publish exactly one half, and the sharpest pairing is a mirror: LangChain prints 1 LCU = $1.50 with no per-action conversion, while Glean prints p50/p90 credit burn for 13 capabilities with no dollar figure anywhere. Six of the ten still carry `price_transparency: public`, which is the point — the synthetic unit is where opacity went after the pricing page stopped being where it lived, and it is invisible to a transparency metric that reads pages rather than equations. Snowflake Cortex and Databricks publish both halves at hyperscaler scale, so this is a disclosure choice rather than a technical limit; if the six-company both-halves cohort keeps growing, the gap closes and this becomes an early-market artefact instead of a structural device.

  LangChain LlamaIndex Mintlify Qodo Vellum GitLab Glean Jasper Photoroom Outreach

  Reviewed August 2026 See the pattern

• The voice minute is unbundling into pass-through component meters

  Sharpens

  9 companies

  Voice-agent platforms are decomposing the conversation minute into a thin per-minute platform fee plus at-cost pass-through of the component stack (LLM, TTS, telephony), making bring-your-own components a first-class discount lever — voice AI's version of the BYOK split between orchestration and inference. Two developments sharpen the claim. First, the bundled counter-camp has begun adopting the mechanic rather than resisting it: Bland AI dropped telephony out of its all-in rate on 2026-07-21, so the rate now covers LLM + STT + TTS only with telephony 'billed separately, on your own carrier or Bland's at pass-through cost.' Telephony is the component every unbundler starts with, because it cannot be made to look like software. Second, the architecture's real weakness is not buyer preference but SALEABILITY: Synthflow published a complete component meter (~$0.11–$0.24/min assembled) and pulled it 15 days later on 2026-06-24, replacing the whole page with a single Enterprise plan from $30,000/year via Contact Sales. It did not re-bundle — it left self-serve, because an unbundled minute is a configuration and configurations need someone to assemble them. Meanwhile the pass-through catalogs keep growing (LiveKit added six per-minute LLM SKUs on 2026-07-21 and its calculator's default total moved from $0.0735 to $0.0672/min) and the components keep deflating (Speechmatics cut real-time enhanced STT and grew its free tier to 50 hours/month on 2026-07-06), which lowers the floor under everyone's assembled minute. The surviving bundlers (Cartesia, Tavus) are now competing on the all-in number rather than defending it on principle — Cartesia deleted its 20%-off annual toggle on 2026-07-22, a ~20–25% effective increase delivered as one opaque number, which is precisely the repricing hazard bundling carries.

  Synthflow Retell AI Vapi Deepgram ElevenLabs Bland AI LiveKit xAI Speechmatics

  Reviewed August 2026 See the pattern

• Localized list prices are making the list price unobservable

  Sharpens

  8 companies

  A small but growing set of vendors is making list price geography-dependent by design — purchasing-power-adjusted local currency, region-restricted serving paths, per-region consumption rates, and even per-region entitlements — and they do not agree on which direction geography should push. Wispr Flow and Synthesia both moved India to rupee list prices on 2026-07-22: Wispr to roughly a third of its US rate, Synthesia off a discounted USD table onto a costlier local one. The enabler is client-side currency resolution, which is also the problem: Wispr's localized price exists only in the rendered page while the served HTML still says $12, so no page states the country rule and a screenshot records whichever price the capture vantage happened to see. This is 7 of 353 companies, not a market norm — OpenAI and Anthropic hold single global USD rate cards and Sarvam AI is INR-native with no USD table at all — but for the vendors that do it, "the list price" has stopped being a single observable number, and our own price verification is blind to it.

  Clipdrop Snowflake Cortex Google Wispr Flow Synthesia Photoroom Fireworks AI Cursor

  Reviewed August 2026 See the pattern

• Outbound sales-tech converges on seat + enrichment credits

  Holds

  8 companies

  Outbound-sales and marketing-automation AI companies have independently converged on the same pricing structure: a per-seat subscription anchoring the base, layered with a rechargeable credit pool. 7 of the 9 outbound/sales-tech vendors in the corpus run it — a 78% sector convergence rate, higher than any other named use-case segment — with only the fully gated, sales-led vendors (11x, Nooks) deviating. As of 2026-07-30 the sharper claim is about what the credit meters. Apollo's first published per-endpoint credit table (2026-07-18) shows it is a shared DATA-consumption balance rather than an AI meter: the API draws down the same pool that funds a rep's in-app email and mobile reveals, and the expensive unit is a data FIELD, not an AI action — people enrichment costs 1 credit for demographics/email but +8 when a mobile phone is returned, a 9x swing decided by what Apollo finds rather than what the caller asked for, with phone waterfalls at 8-25 credits (45+ in some configs) and create/update/list endpoints at 0. So the seat is the stable denominator (the rep) and the credit is the variable numerator, but the numerator is driven by data acquisition cost, not inference cost — which is why this segment's credits behave differently from the compute-metering credits in AI coding and creative tools. The pool is also where repricing lands in both directions: Apollo halved included credits in 2025 without touching dollar prices, Smartlead raised all four tiers while cutting Prime's send cap 5.69M → 500K/mo (2026-07-14), and Juicebox cut its Starter seat 14% while doubling bundled credits and lifting Growth's allowance 50% (2026-07-21). Budget the seat and the pool separately, and ask which data field triggers the expensive rate.

  Apollo.io Lemlist Reply.io Artisan Regie.ai Smartlead Instantly Juicebox

  Reviewed August 2026 See the pattern

• Search APIs split: commodity calls deflate, 'research effort' inflates

  Sharpens

  8 companies

  Agent-facing search APIs have bifurcated into two price layers moving in opposite directions, and by July 2026 the commodity layer reached zero while the effort layer acquired published ceilings. The commodity per-call layer converged around $5–7 per 1,000 requests and then went free at the margin: You.com added 100 free Web Search calls per DAY before its $5/1k rate applies (2026-07-29), Exa abandoned request caps entirely for $20 signup plus $10/month of credits with no payment method required (2026-07-29), and Firecrawl bundled an identical +1,000 search credits into every plan from Free to Enterprise (2026-07-30). The effort layer is where margin is rebuilt, and it is now priced explicitly rather than gestured at: You.com replaced '>$2,000/1k calls' on its Frontier research tier with a firm $1,200/1k on 2026-07-22 — a 40% cut that also produces a published 240× spread against its own $5/1k search — while Exa halved its Agent ceiling ($0.025–$2.00 → $0.012–$1.00 per run) and added a Minimal mode at the bottom, giving five fixed-effort modes spanning 83×. A new effort UNIT also appeared: Linkup's closed-beta Extract endpoint bills ~$2–$10 per TASK, not per call. By August 2026 the commodity layer had also broken through its own floor and absorbed synthesis: Perplexity's Agent API web_search settled at $0.0025 per call ($2.50/1k, half the band) while it REVERSED the identical cut to fetch_url on 2026-08-14, and You.com launched an Answer API that runs search plus citation-grounded synthesis in one call at $5.00/1k — exactly its raw-search rate — despite already operating a $12-to-$1,200/1k Research ladder. Reasoning alone no longer earns the effort premium; only multi-step research does. Diffbot (2026-08-05) confirms the floor from outside: it added Web Search at 1 credit per query, the same as a page extraction, bundled free into every tier. Vendors that cannot sell effort (SerpApi, proxy networks) stay on pure volume ladders and compete on deflation alone — Oxylabs restructured its Web Scraper API page twice in eight days with every per-1K rate byte-identical. The buyer implication has inverted: retrieval and now single-shot synthesis are becoming free and unpriceable, so the entire negotiation is about multi-step effort — and effort ceilings are still moving 40% in a week.

  You.com Exa Tavily Linkup Groq Firecrawl Perplexity AI Diffbot

  Reviewed August 2026 See the pattern

• Support and security entitlements climb the ladder

  Holds

  8 companies

  A price rise that never appears as one: 8 of 380 corpus vendors held every published seat rate constant while shrinking what the seat entitles you to — email support, SSO, SCIM, audit-adjacent governance, SLAs, retention policy, bundled professional-services hours all moved up a tier or out into an add-on. The enabler is that these entitlements live in a feature-comparison table rather than on a price card, so the change is invisible to any tracker watching dollar figures, and in LangSmith's case invisible even on the plan card that used to advertise it. Perplexity (2026-08-14) is the first case at a PUBLISHED per-seat price: it brought its enterprise pricing page back online at $34 and $271 per seat/month — consistent with the prior annual totals, so not a repricing — while moving SCIM, audit logs, RBAC, data-retention configurability and a separate enterprise domain to "Contract required" on BOTH tiers, replacing a 50-seat threshold that used to unlock them. Fireworks (2026-08-11) priced a compliance entitlement for the first time: US-only or Europe-only pinned deployments now cost a flat 1.5x the standard on-demand rate behind a Contact Sales request. The boundary condition is that the direction is genuinely unsettled — 5 vendors have published the same entitlement class DOWNWARD, one (Lago) ran both directions in a single change, and the largest single downward move came AFTER this signal was first logged: Freepik/Magnific (2026-08-06) moved five governance and support entitlements — centralized admin control, no-training-on-your-data, Magnific Studios access, dedicated-CSM onboarding, metadata export/lineage — from Enterprise-only onto the $69/seat Business card at an unchanged price. So this may still be governance competition rather than governance rationing. The open risk is that it only becomes visible at renewal, which is also when it becomes a churn event.

  LangSmith Fathom Lago CrewAI Apify Zendesk Perplexity AI Fireworks AI

  Reviewed August 2026 See the pattern

• The unit rate goes dark while the tier price stays public

  Weakens

  8 companies

  Fourteen corpus vendors have withdrawn a previously published per-unit, overage or quantity figure while leaving a headline price on the page, so "is pricing public?" has stopped being a binary question — 5 of the original 8 still carry `price_transparency: public`, and the corpus's own top-line metric is blind to the move (108 of 380 companies, 28%, are gated or sales-only, a rate that has now FALLEN for three consecutive reviews — 30%, 29%, 28% — even as these withdrawals landed). The enabler is that the tier fee is the cheap, stable half of an AI bill and the marginal rate is the negotiable half, so pulling the rate costs a vendor almost no self-serve conversion while moving the entire variance of the bill into a sales conversation. BUT THE DIRECTION IS INVERTING, and this trend published the condition under which that matters: "if the disclosure cohort keeps growing faster, the direction inverts." In the 2026-07-31 to 2026-08-16 window it did — 7 new disclosures against 6 new withdrawals, taking the cumulative score from 8-vs-6 to 14-vs-13, essentially parity. Two of the vendors originally cited as counterexamples now sit on both sides: Metronome deleted the included-volume quantities from the same Starter card it had published three weeks earlier (the rates survived, the quantity did not), and Modal, cited for itemising two hidden multipliers, gated a token-billed Shared API it has still never published rates for. The durable claim is narrower than the original: withdrawal is real, it is invisible to the transparency field, and it targets QUANTITIES at least as often as rates — but it is not winning, and it should not be read as a ratchet.

  Arize AI CrewAI Resemble AI Synthflow Usage AI Tavus Zhipu AI Braintrust

  Reviewed August 2026 See the pattern

• Zero-markup resale transmits a model cut in days

  New

  7 companies

  A layer of the corpus no longer sets the price of the models it sells. Five companies — Glean, Cursor, Augment Code, LiveKit and Vercel — moved the same two OpenAI SKUs by the same two percentages inside eleven days without announcing it, because each bills third-party model tokens at cost inside a pooled meter. Netlify shows the same mechanic on a different upstream, halving Gemini 3.6 Flash at a fixed 180-credits-per-dollar conversion. The boundary condition is the zero-markup policy itself: a vendor that sets its own hosting margin does not transmit, and the three counterexamples all move rates their own way, one of them upward. The buyer consequence is that a coding-assistant bill is a derivative of a rate card the buyer never sees, and the vendor's own change-log entry is not a pricing decision. Falsified if a zero-markup reseller holds a rate after its upstream provider moves, or if the same synchrony appears among vendors that do take model margin.

  OpenAI Glean Cursor Augment Code LiveKit Vercel Netlify

  Reviewed August 2026 See the pattern

• Homegrown billing gets retired as companies scale — and sometimes the meter goes with it

  Holds

  6 companies

  Some usage-metered AI companies start with a hand-rolled billing system and retire it for a vendor once the product line and scale exceed what the homegrown version supports. Three clean arcs now exist: Pinecone (homegrown engine that still required manual invoices, replaced by Orb as system of record at serverless launch), Langfuse (homegrown per-event billing replaced by Stripe Billing/Checkout/Tax), and Replicate (a homegrown metering-and-fraud stack torn out for Metronome, prepaid credits shipped in 14 days, 2025-09-01). The originally-claimed boundary — that only the invoicing/plan layer moves while the meter stays in-house — holds for Pinecone and Langfuse but NOT for Replicate, which handed the meter over too and said why: fraud durability was worth more than owning the meter. Two adjacent variants exist: a vendor→vendor arc (Yellow.ai kept Zuora for subscriptions, hit 'performance and feature limits at high usage' on its metered layer, and layered Togai on top) and a rev-rec/ERP analogue (LiveKit QuickBooks→NetSuite; Outreach NetSuite→SAP plus a from-scratch CPQ replacement, 2026-06-29). The pattern is real but distinctly a minority path: of the 19 corpus companies running the clean in-house-meter + bought-biller split, only 3 disclose retiring a prior homegrown biller — the rest designed it that way from the start (Chroma, Vercel), skipped the homegrown phase entirely (Browserbase, Mistral 'instead of building a billing system from scratch'), or run the counter-direction and build at scale (Synthesia's in-house biller kept through $40M→$140M ARR on ~one engineer; Anthropic's homegrown ledger; Mercor designing an immutable ledger 'from scratch' to move $200M). The trigger, where the migration does happen, is consistent: multi-product billing, custom plans and version-controlled pricing that the hand-rolled system could not deliver without becoming an engineering bottleneck — or, in Replicate's case, a fraud-and-maintenance burden.

  Pinecone Langfuse LiveKit Replicate Yellow.ai Outreach

  Reviewed August 2026 See the pattern

• SEO tools reprice around AI-search visibility

  Holds

  6 companies

  A cohort of five SEO and AI-writing tools has repriced around generative-engine optimization (GEO): instead of selling words or documents, they meter and gate on AI-search visibility — 'visibility prompts' tracked and the number of answer engines (ChatGPT, Perplexity, Google AI Overviews, Gemini, Claude) monitored. The migration completed three further steps by 2026-07-30. (1) The value metric graduated from a bundled allotment to a STANDALONE product: Surfer SEO now sells AI Search Analytics on its own pricing tab, metered purely by a prompts-tracked-per-DAY slider ($82 / $158 / $308 per month annual for 50 / 100 / 200 prompts a day, unlimited via sales), beside an unchanged $49–$299 content-optimization ladder — nine months after the same capability was a $95/mo add-on for 25+ prompts on a weekly refresh. (2) The cohort's flagship holdout converted: Jasper, previously cited here as the per-seat incumbent with 'no meter of any kind', published a public rate card on 2026-07-22 charging 10 credits per GEO Hub query or page run, 10 per Grid row, 40 per Research/Translation Agent run and 100 per Optimization Agent run — metering reintroduced three years after Jasper killed its word caps, specifically to price AI-search work, with no dollar-per-credit rate published. (3) The meter multiplied: Frase now bills seven dimensions at the same $49/$129/$299 prices, makes per-tier AI-engine coverage an explicit entitlement, and ships opt-in-only overage that is off by default. The direction of travel is clear — engine coverage and prompt volume are becoming the ladder — but the shift is still concentrated in tools whose buyers already lived and died by search rankings. Rytr still meters characters, and localization tools (Crowdin, Lokalise) meter seats and words.

  Writesonic Frase Surfer SEO Scalenut Apollo.io Jasper

  Reviewed August 2026 See the pattern

• IPO track ends the discount era

  Weakens

  3 companies

  A listing event may mark the end of a lab's promotional-pricing phase: Zhipu AI removed first-purchase discounts 34 days after its 2026-01-08 Hong Kong IPO and raised the GLM Coding Plan list 80% at 195 days. The proposed enabler is margin disclosure — a listed lab has to defend gross margin quarterly, which makes a first-purchase promo expensive in a way a private lab does not feel. This is a hypothesis on n=1 complete, not a pattern: Moonshot AI ran both price moves before its listing process surfaced, 01.AI moved new revenue off its published card without touching it, and Anthropic widened its funnel 13 days after IPO-preparation reporting. The corpus-wide version of this idea has already failed — funding events do not predict repricing here — so only the narrower listing sub-case has any support at all.

  Zhipu AI Moonshot AI 01.AI

  Reviewed August 2026 See the pattern

• The middle of the ladder churns — but the free tier is not the stable anchor this trend assumed

  Superseded

  17 companies

  NARROWED at the 2026-07-30 re-test after the trend's own kill condition fired. What SURVIVES: the middle self-serve band — roughly $10–$200/mo Starter/Pro/Plus tiers — is still the single most-repriced part of an AI pricing ladder, getting re-priced, re-metered, split, renamed or deleted more often than any other band (about 15 mid-band edits in the 2026-07-15 to 2026-07-30 window), and its two terminal forms remain observable: fragmenting (WellSaid split one $50 tier into $10 + $33) or deletion (Synthflow removed self-serve entirely for a $30K/yr enterprise floor; Qodo deleted per-seat Teams for a pooled credit balance). What is FALSIFIED: the mechanism. The free tier is not a pinned top-of-funnel acquisition gate — in that same 15-day window free-tier terms moved 12 times, 8 tightenings or deletions (cerebras open Free → one-time $5 credit trial; freshworks deleted its $0 plan; gladia swapped a recurring 10-hours/month allowance for a one-time 50 EUR credit; photoroom removed its Free card; maxio dropped its free sandbox; playground cut Free downloads 10→2/day; gumloop killed a permanent $0/5,000-credit plan for a 14-day trial; puzzle replaced a volume-gated free tier with a 2-month promo) and 4 expansions (socket 10x'd its free dependency cap to 10,000; exa doubled the signup credit to $20 and dropped the card requirement; you-com ADDED a 100-calls/day free tier; firecrawl bundled +1,000 search credits into every tier including Free). The enterprise end moved 7 times too (inflection-ai retired its Enterprise offering to a 404; runpod launched an Enterprise agreement surface; fathom added a quote-only Enterprise tier; langsmith moved Email Support to Enterprise-only; lago collapsed Business and Enterprise into one Premium; abridge named an unlimited enterprise agreement; 01-ai launched a quoted enterprise platform). So instability is GENERAL, merely concentrated in the middle — roughly 15 / 12 / 7 across mid / free / enterprise. The correct reading is that the free tier has become an active monetization lever in both directions rather than a fixed acquisition constant, and buyers should treat free-tier terms as no more durable than a mid-tier price. Counterexamples to the mid-band concentration itself remain (E2B's $150 Pro held identical from 2024-12 to 2026-05; Intercom's Advanced $85 and Expert $132 held while its ENTRY tier moved), so even the surviving claim is a tendency.

  WellSaid Dust Runway Tavus Windsurf Synthesia Ideogram Galileo Synthflow Shortwave Creatify Qodo Lokalise Mem0 Freepik Vellum Poe

  Reviewed August 2026 See the pattern

• Agents priced as headcount: the salary anchor enters the rate card

  Superseded

  6 companies

  SUPERSEDED 2026-08-16. Pricing agents as headcount — named workers at flat monthly salaries anchored explicitly against human payroll — is dead at the published end of the corpus. As of Lindy's 2026-08-11 restructure, ZERO of the 380 corpus companies publish a payroll anchor on a rate card. Three of the seven vendors this trend cited have now abandoned the frame, and all three replaced headcount with a conventional unit: Motion reverted its 'AI Employees' grid ($29–$599/mo) to credit plans by December 2025; Artisan deleted the corpus's only seniority-named ladder (Intern $250, Employee $600) on 2026-07-14 for quote-only tiers sized in leads contacted per month; and Lindy — the vendor whose struck-through '$8,000/month human assistant' column was this trend's entire argument — deleted the anchor on 2026-08-11 and moved to per-seat billing with a published credit meter (Plus cut 40% to $29.99/user/mo with 3,000 pooled credits). AiSDR, in the category that invented the digital-worker pitch, moved to AI-researched-contacts metering on 2026-06-20 as its third metering model in under two years. The surviving-vendor count fell 5 to 4 (11x, Reply.io, Juicebox, Dropzone AI) — the first decline this trend has recorded — and no new adopter has appeared since 2026-06-04 across a corpus that grew 353 to 380. What remains is a narrower and different claim: a flat per-agent line ITEM sold beside a conventional seat or metered plan, with no payroll comparison attached, at four vendors of which two publish no price at all. The persona survives as marketing; the meter underneath won.

  Artisan 11x Lindy Reply.io Juicebox Dropzone AI

  Reviewed August 2026 See the pattern

• Annual prepay discount compression — a repricing tactic, not a category direction

  Superseded

  3 companies

  Some self-serve AI vendors remove or gut the annual-prepay discount as part of a price increase, and the tactic has a consistent shape: they DELETE the incentive rather than tune it. Three clean adopters: Creatify cut annual billing from 'up to 50%' to 'up to 15%' in the same change that doubled Pro $49 to $99 (2026-06-30); Shortwave deleted the annual/monthly toggle entirely as it raised every paid tier 20-29% (2026-06-30); Cartesia removed a Monthly/Yearly toggle whose Yearly view was both badged 'Save 20%' and set as the DEFAULT, so its advertised prices went from $4/$39/$239 to $5/$49/$299 — a 25% effective list rise with no committed-term alternative left on the page, and with every credit allowance and per-minute rate unchanged (2026-07-22). The mechanic is that a deep annual discount lets a buyer lock the old cheap rate a year forward, so a vendor repricing upward stops pre-selling it. The important negative finding is that this is a TACTIC, not a category direction: in the same 2026-07-15 to 2026-07-30 window five vendors moved the other way. VEED raised every seat 5-20% while keeping Yearly selected by default; Fathom DEEPENED its annual discount from 'save 22%+' to 'save 25%+'; Resemble INTRODUCED a 20% annual saving on two new tiers; Photoroom pruned its WEEKLY option and kept Yearly at 'Save 33%'; Gumloop gated only its free TRIAL to monthly while leaving '20% off annually' on paid plans. Cash-flow and churn economics keep pulling the other way, and they are winning on volume. The buyer-relevant form of the claim is diagnostic rather than predictive: a vendor shrinking or deleting its annual toggle is very likely mid-reprice, and the toggle removal itself can be the price increase — Cartesia's headline numbers 'rose' 25% purely because the default view flipped from yearly-equivalent to monthly.

  Creatify Shortwave Cartesia

  Reviewed August 2026 See the pattern

No trends in this view.