The unit rate goes dark while the tier price stays public
Fourteen corpus vendors have withdrawn a published per-unit, overage or quantity figure while leaving a headline price on the page — so 'is pricing public?'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. But the direction is inverting: the trend published a falsifier, and it fired. Withdrawals and new disclosures now sit at 14 versus 13, essentially parity.
What's happening — and why
What's happening: vendors pull the marginal rate — the overage price, the per-unit cost, the included quantity — while the tier fee stays published. Five of the eight originally cited still carry price_transparency: public, so the corpus's own top-line metric cannot see the move at all.
Why: the tier fee is the cheap, stable half of an AI bill and the marginal rate is the negotiable half. Pulling the rate costs a vendor almost no self-serve conversion while moving the entire variance of the bill into a sales conversation.
Why the status is 'weakens': this trend published the exact condition under which it would stop being true — 'if the disclosure cohort keeps growing faster, the direction inverts' — and in the following window it did, with 7 new disclosures against 6 new withdrawals, taking the cumulative score from 8-vs-6 to 14-vs-13.
The durable claim is therefore 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.
How it works
Evidence over time
8 supporting · 13 counter — hover or tap a point for detail, click to jump to the row.
Evidence
| Company | Date | What happened |
|---|---|---|
| Arize AI | Jul 2026 | The cleanest case in the corpus. Headline prices (AX Free $0, AX Pro $50/mo, AX Enterprise custom) and every included quota (50,000 spans, 10 GB ingestion, 30-day retention on Pro) are unchanged, but the previously published $10-per-million-trace-spans and $3-per-GB overage rows were removed when the page was rebuilt into a collapsible feature-comparison table on the same date. A re-capture with every accordion section expanded confirms the figures no longer appear anywhere on the page. Arize still records `price_transparency: public`. |
| CrewAI | Jun 2026 | The $0.50-per-execution overage rate vanished between the 2026-05-03 and 2026-06-03 snapshots and was replaced by the phrase "flexible overage." The $25 Professional tier had already been deleted by 2026-05-03, so the public page's only remaining number is "Free" — Basic at 50 workflow executions/month with a hard cap and no overage path — plus a custom Enterprise column whose included allotment is "sized to workflow." A full transparency reversal: public three-tier pricing with a $0.50 unit rate (Oct 2025) to mid-tier deleted (spring 2026) to unit rate deleted (mid-2026). |
| Resemble AI | Jul 2026 | Every voice-generation usage rate was removed from the pricing page — text-to-speech $0.0005/sec, voice agents and speech-to-text $0.001/sec, audio enhancement $0.002/sec, audio editing $0.0005/sec — and the generative-voice products were regrouped under a non-commercial "Voice AI Research" footer heading. Detection, verification and watermarking rates were retained, so the page still reads as a public rate card. The Enterprise discount claim also softened from "volume discounts up to 80%" to "volume discounts based on commitment," and the sales nudge rose from $500/mo to $2,000/mo. |
| Synthflow | Jun 2026 | The composable per-minute rate card — $0.09/min Voice Engine plus LLM $0.02–$0.05/min plus telephony $0.00–$0.02/min, roughly $0.11–$0.24 all-in with 5 concurrent calls included — was removed 15 days after being captured live on 2026-06-09. The page now shows a single Enterprise card whose only published number is a $30,000/year contract floor, with call volume, concurrency, telephony, integrations and security all scoped via Contact Sales. The PAYG terms are grandfathered for existing customers only. |
| Usage AI | Jul 2026 | The pricing page's inline fee calculator (a worked example putting a $1,200 fee on $8,000/mo of realized savings, a ~15% effective take) and its Insured-Commitment-Rate-vs-Effective- Savings-Rate comparison table were both deleted. The fee is now described only as "a percentage of the savings we realize for you, billed monthly against verified savings," with no number anywhere on usage.ai. Because the model is pure savings-share, the withdrawn percentage IS the price; the 20% (EC2) / 35% (RDS, ElastiCache, Redshift, OpenSearch) rates survive only on the company's AWS Marketplace listing. |
| Tavus | Jul 2026 | `tavus.io/pals/pricing` began 301-redirecting to the developer pricing page, so the consumer PALs ladder (Free / Plus / Max, gated at 15 / 150 / 500 call minutes) is no longer published on any Tavus surface — the product still ships, with an iOS app and a PAL Maker, but the page carries no dollar figure. The same capture relabelled the $22 Starter card "No pay as you go coverage," making the entry developer tier a hard cap with no published metered path above it. Developer per-minute overages were untouched ($0.35 Builder / $0.31 Growth / $0.26 Business). |
| Zhipu AI | Jul 2026 | The quantity-side variant, executed alongside a price rise. GLM Coding Plan list prices went to $18 / $72 / $160 per month (from roughly $10 / $30 / $80, a ~26% rise at Lite and ~40% at Max) while the published quota changed from an absolute "~80 prompts per 5 hours" to relative multiples — "Pro = 5x Lite usage, Max = 20x Lite usage." Cross-tier comparison got easier and absolute capacity became unknowable, so no published number now converts $18 into work done. |
| Braintrust | Jul 2026 | The mildest form, and the one to watch because it is the cheapest to execute. The pricing page stopped printing its per-tier "$0.06/mtok input, $0.40/mtok output" line on the plan cards, replacing it with a generic "then token rates" plus a "View detailed pricing, including model and token rates" link. No rate changed — Topics overage holds at $0.06/$0.40 per mtok, GLM-5.2 at $1.40 in / $0.26 cached / $4.40 out, and the $0 / $249 / custom tier fees are untouched. The rate moved one click away rather than disappearing, and the credit pool was renamed from "Topics credit" to "Model credits" in the same pass. |
Counterexamples
- Glean · Jul 2026 — Expanded disclosure hard in the same month and in the same direction this trend says is closing. Enterprise Flex docs went from a short query-type list to a 13-line FlexCredit rate card with p50/p90 consumption per capability (Adaptive Reasoning ~11/~38 standard and ~26/~83 premium, Voice Session ~3/~26, Meeting Notes ~9/~18 per minute, Deep Research ~33/~144, Advanced Agent Runs ~150/~450). On 2026-07-21 it published Glean Core Suite with per-million-token dollar rates — the first dollar figures Glean has ever exposed publicly — and on 2026-07-15 shipped hard, configurable spend caps.
- Apollo · Jul 2026 — The exact opposite move: per-endpoint API credit rates that had lived behind a login-gated "About Credits" page were published for the first time — 1–9 credits per enriched person (1 for demographics/email, +8 if a mobile phone is returned), email waterfall 1–4, phone waterfall 8–25, and 0 credits for create/update/list/manage endpoints. No rate moved; a gated rate simply became public.
- Metronome · Jul 2026 — Reversed a years-long no-published-rates posture: the Starter plan now shows $100,000 of billing volume and 10M events included per month, then 0.8% of billing volume and $0.04 per 1,000 events. Custom remains sales-quoted. A billing vendor that published none of its own prices printed a marginal rate.
- Modal · Jul 2026 — Turned two previously implicit multipliers into published line items: region selection at 1.5–1.75x base prices and non-preemptible (guaranteed) execution at 3x base prices, itemised on the pricing page for the first time across Starter, Team and Enterprise. All existing per-second GPU/CPU/memory rates and plan fees held, so the rate card gained resolution without a repricing.
- Lorikeet · Jul 2026 — Published its whole rate card as machine-readable Markdown at `lorikeetcx.ai/pricing.md`, linked from the plan cards as "Pricing also available in machine-readable format" — plan table, per-ticket credit rates and worked billing examples, parseable by a procurement script. The same capture cut the other way in miniature, though: the voice-resolution credit rate (1.50 credits on Start, 1.20 on Scale) gained a footnote scoping it to calls up to 3 minutes, with no published rate for anything longer.
- Jasper · Jul 2026 — Published a standing public rate card (stamped "Updated 6.9.2026") pricing Grid rows and GEO Hub runs at 10 credits, Research and Translation Agents at 40, the Optimization Agent at 100, and API/MCP calls at 1–40 — but never a dollar per credit. A disclosure move that supplies only one half of the equation, which is why Jasper is evidence in the sibling trend `credit-equation-published-in-halves` rather than here.
- GitLab · Jun 2026 — The inverse withdrawal, and the reason this trend is about selective withdrawal rather than overage rates specifically. GitLab pulled a long-standing public $29/user/month Premium list price behind "Let's talk" (Ultimate had already gone to "Contact us" by mid-2024) while keeping all three of its unit rates published: GitLab Credits at $1/credit, CI/CD compute at $10 per 1,000 minutes, storage at $5/month per 10 GiB. The tier price went dark and the marginal rate did not — on a NASDAQ-listed vendor.
- Groq · Aug 2026 — The best-fitting new adopter this trend has, because the withdrawal is surgical. groq.com/pricing now 308-redirects to a homepage carrying no pricing content and no Pricing nav link — and every per-token, per-hour and per-character rate survives byte-identical inside the GroqCloud developer docs (Llama 3.1 8B $0.05/$0.08, Llama 3.3 70B $0.59/$0.79, GPT OSS 120B $0.15/$0.60, GPT OSS 20B $0.075/$0.30, Whisper $0.111/$0.04 per hour, Orpheus $22/$40 per 1M characters), unchanged from the 2026-07-21 capture. Exactly three SKUs went dark: the built-in agentic tools for web search, website visits and code execution, whose docs defer to the pricing page that no longer exists. A subset of marginal rates disappeared while the rest stayed public.
- WellSaid · Aug 2026 — A new shape: the meter arrives WITHOUT the rate, rather than the rate being withdrawn. WellSaid's Starter and Pro cards each gained an "Extra minutes available" line — the first self-serve overage mechanism the company has published, addressing a long-standing no-overage gap — with no published per-minute rate anywhere. The same capture converted the 7-day Trial into an evergreen free plan with 3 downloaded minutes/month (no commercial rights). A buyer can now exceed their allowance self-serve and cannot compute what it costs.
- Metronome · Aug 2026 — A counterexample partially reversing itself in three weeks, and it withdrew a QUANTITY rather than a rate. From 2026-07-14 through at least 2026-07-22 the Starter card read "$100,000 in billing volume included, billed at 0.8% after that" and "10M events included, billed at $0.04 / 1k events after that." As of 2026-08-04 the same bullets read only "Billing volume billed at 0.8%" and "Events volume billed at $0.04/1k ingest events." The metered rates are unchanged; the included allotment is simply gone from the page. This is the Zhipu mechanic (absolute quantity replaced by nothing) at the vendor this trend cited as its cleanest disclosure — and it means a buyer can still compute the marginal rate but no longer knows where metering begins.
- Modal · Aug 2026 — The other counterexample cutting the other way, on a NEW SKU. Modal's July 29 announcement of an OpenAI-compatible, token-billed Shared API described it as covered by Starter's $30/month free-compute offer, implying access on any plan; the same post now says token-based pricing is for Team ($250/mo + compute) and Enterprise only, with Starter routed to a per-second Auto Endpoint for the same models. Modal has not published per-token input/output rates for the Shared API on its pricing page or its billing docs at any point. A vendor previously credited here for itemising two hidden multipliers launched a metered product and never printed its meter.
- Netlify · Aug 2026 — A clean second instance of the GitLab inverse, and the reason this trend is about SELECTIVE withdrawal. Netlify deleted the "starts at $500/month" floor caption beneath its Enterprise "Custom" label between the 2026-07-29 and 2026-08-04 captures — the tier price went dark — while every unit rate stayed published and was verified unchanged in the same capture: 15 credits per production deploy, 10 credits per GB-hour of compute, 20 credits per GB of bandwidth, 2 credits per 10,000 web requests, and 180 credits per $1 of underlying model spend, alongside Free $0, Personal $9 and Pro $20 base with 3,000 credits. Tier price withdrawn, marginal rates intact.
- Bright Data · Jul 2026 — Bounds the pattern: deleting a published rate row is not automatically an opacity move. Web Unlocker, SERP API and Web Scraper API were restructured from a four-tier committed ladder to Free / PAYG / Scale / Enterprise, removing the $1.1 and $1.0 per-1k committed rungs — but $1.5/1k PAYG and a $1.3/1k Scale overage remain published, so the marginal rate is still knowable. What was lost was granular volume discounting, traded for a recurring 5,000-results/month card-free free tier.
Trivia
-
Five of the eight vendors in this trend still carry `price_transparency: public` in the corpus — Arize (2026-07-29), Resemble AI (2026-07-14), Tavus (2026-07-21), Braintrust (2026-07-30) and Zhipu (2026-07-22). Arize is the extreme case: the $50/month AX Pro price, the 50,000-span and 10 GB included quotas and the 30-day retention are all still printed on the page, and a re-capture with every accordion section expanded confirmed the $10-per-million-span and $3-per-GB rows are simply not there. A page can be fully public and still leave the marginal cost unknowable.
-
Resemble AI proves the tier price is the reversible half and the unit rate is not. On 2026-07-14 it stripped every voice-generation per-second rate from its pricing page and moved the products into a non-commercial "Voice AI Research" grouping. Fifteen days later, on 2026-07-29, it reintroduced two subscription tiers above Flex — Team at $280/mo annual ($350 monthly) and Business at $800/mo annual ($1,000 monthly) — and even recut a retained detection rate from $0.04/sec to $0.035/sec. The tier prices came back; the generation rates did not.
-
Tavus runs both postures at once on the same value metric. As of 2026-07-21 its developer Conversational Video Interface plans publish per-minute overages down to three tiers — $0.35/min on Builder, $0.31 on Growth, $0.26 on Business — while `tavus.io/pals/pricing` 301-redirects to that developer page and the consumer PALs ladder (Free / Plus / Max, gated at 15 / 150 / 500 call minutes) is no longer published anywhere on the site. Same company, same conversation minute; the only variable is which buyer is looking.
For buyers
Get the marginal rate in the contract, not the tier price. The tier fee is the part you can already see and the part least likely to hurt you; the overage rate and the included quantity are where the variance lives, and they are exactly what gets withdrawn. Watch quantities as closely as rates — Metronome deleted the included-volume quantities from the same Starter card it had published three weeks earlier, and the rates survived while the quantity did not. If a vendor will not put a marginal rate on paper, price the deal as if that rate is unbounded, because contractually it is.
For vendors
Withdrawing the marginal rate is cheap in conversion terms and expensive in trust terms, and the asymmetry only shows up at renewal. If your marginal rate genuinely has to be negotiated — because it depends on volume, region or commitment — publish a band rather than nothing. A published range preserves almost all of your negotiating room and removes the impression that the number is being hidden because it is bad. Note also that the market is not moving your way on this: disclosures now roughly match withdrawals across the corpus.
Outlook — what to watch
The trend is at parity and should be re-tested rather than restated. Two of the vendors originally cited as counterexamples now sit on both sides, which is the clearest sign that this is oscillation rather than direction — Metronome deleted included quantities from a card it had just published, and Modal gated a token-billed API whose rates it has still never published. Watch whether the disclosure side keeps its lead; if it does for another cycle, the trend should be superseded by the broader re-sorting story in floors-go-dark-burn-rates-go-public, which explains both directions at once.
Bottom line
Fourteen vendors pulled a marginal rate while keeping a headline price, and the corpus's transparency field recorded none of it. But the direction has inverted to parity against the trend's own published falsifier, so read this as a real mechanic that is not winning.
FAQ
What does it mean for a unit rate to go dark?
The vendor keeps the tier price on the page — $499 a month, say — but removes the overage rate, the per-unit cost or the included quantity. You can still see what the plan costs and no longer see what exceeding it costs, which is where the variance of an AI bill actually lives.
Why doesn't this show up as reduced transparency?
Because the corpus field price_transparency reads pages, not equations. Five of the eight originally cited vendors still carry public, since their pricing pages do show a number. Corpus-wide, gated-or-sales-only has actually fallen for three consecutive reviews even as these withdrawals landed.
Is this trend still true?
Partly. It published a falsifier — if the disclosure cohort grows faster, the direction inverts — and the falsifier fired: 7 new disclosures against 6 new withdrawals took the cumulative score to 14 versus 13, essentially parity. Withdrawal is real and invisible to our metric, but it is not winning, and it should not be described as a ratchet.
What should I insist on contractually?
The marginal rate and the included quantity, both in writing. Withdrawal targets quantities at least as often as rates: Metronome removed the included US$100,000 billing-volume and 10-million-event allotment from its Starter card three weeks after publishing it, leaving the 0.8% and $0.04-per-1,000-event rates to apply from the first dollar.