The credit equation is published in halves
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. Ten corpus companies publish exactly one half: either dollars-per-unit or units-per-action, almost never both. Six of the ten still carry price_transparency: public.
What's happening — and why
What's happening: to know what an action costs you need two numbers — how many dollars a credit is worth, and how many credits an action burns. Vendors reliably publish one and withhold the other, which makes the synthetic unit an opacity device rather than the simplification it is sold as.
The sharpest illustration is a mirror pair. LangChain prints 1 LCU = $1.50 with no per-action conversion anywhere. Glean prints p50 and p90 credit burn for 13 separate capabilities with no dollar figure anywhere. Each publishes exactly what the other withholds, and neither lets you compute a cost per action.
Why this matters more than it sounds: six of the ten still carry price_transparency: public in our corpus, because their pricing pages do show numbers. The synthetic unit is where opacity went after the pricing page stopped being where it lived, and it is invisible to any transparency metric that reads pages rather than equations.
This is a disclosure choice, not a technical limit — Snowflake Cortex and Databricks publish both halves at hyperscaler scale.
How it works
Evidence over time
14 supporting · 9 counter — hover or tap a point for detail, click to jump to the row.
Evidence
| Company | Date | What happened |
|---|---|---|
| LangChain | Jul 2026 | Collapsed seven per-service meters into two normalized currencies with exact dollar values — 1 LCU = $1.50 (work and compute: Engine, Fleet, deployment runtime, sandboxes) and 1 LSU = $1.00 (traces and storage). The consumption side is published only in infrastructure terms: 0.045 LCU/vCPU-hr runtime compute, 0.006 LCU/GiB-hr runtime memory, 0.177 LSU/vCPU-hr database compute, 0.025 LSU/GiB-hr database memory. The one per-action figure published is a band — an Engine run consumes "roughly 5–30 LCUs" and Engine fires every 6 hours — so a buyer must model vCPU-hours to reach a price, and the band alone spans 6x ($7.50 to $45 per run). Seats held at $0 / $39 / custom. |
| LlamaIndex | Jun 2026 | Prepaid credit bundles at an exact $1.25 per 1,000 credits in both US and EU regions, across Free (10K credits/mo), Starter $50/mo (40K) and Pro $500/mo (400K). The per-parse consumption rate — how many credits a document actually costs to index — is not on the plan card. The dollar half was repriced from $1.00 to $1.25 per 1,000 in February 2026 with included credits trimmed ~20% and headline plan prices held, making each paid plan's price exactly equal to the face value of its included credits. |
| Mintlify | Jun 2026 | Free Starter plan with 5,000 AI credits included and an exact $0.01/credit overage, plus a quoted Enterprise tier and a buyer-settable hard cap. Which action costs how many credits is not published: the assistant, the writing agent and workflows all draw on the same pool with no per-feature rate, so the $0.01 anchor cannot be converted into a cost per documentation page or per assistant query. |
| Qodo | Jun 2026 | Pooled team-wide credits at an exact $0.012 each with no per-seat fee, sold as selectable packs of 2,500 / 5,000 / 20,000. The conversion is published only as an estimate range — the cards label those packs "~18 / ~36 / ~144 reviews per month" — so cost per code review is a band, not a rate, and the pack labels are the only bridge between the dollar rate and the work. Premium-model multipliers exist (Claude Opus at 5 credits, Grok at 4 in the prior seat-based packaging) but no current per-action table is published. |
| Vellum | Jul 2026 | The degenerate dollars-only case: "$1 = 1 credit," passed through to model providers at cost with no markup, so the dollar identity is exact by construction — and precisely because of that Vellum publishes no per-call credit draw. The conversion is whichever provider you routed to at whatever their token price is that week. Vellum monetizes machine tiers (Medium +$35, Large +$60, XL +$125/mo), storage and a $10 platform fee instead, and preset Pro packages bundle credit allowances (Mighty $25, Super $45, Ultra $115). |
| GitLab | Feb 2026 | GitLab Duo Agent Platform went GA with GitLab Credits at a clean $1/credit, with Premium bundling $12 and Ultimate $24 of monthly credits per user for a limited time. No credits-per-agent-action rate is published, so the anchor is unusable for forecasting: a buyer knows the currency's exact value and nothing about how fast an agent spends it. Note GitLab is a counterexample in the sibling trend `the-unit-rate-goes-dark` for the opposite reason — four months later it pulled the $29 Premium seat price behind "Let's talk" while keeping this $1/credit rate published. |
| Glean | Jul 2026 | The mirror image of LangChain, published one day later. Enterprise Flex docs carry 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, Code Writer ~9/~32, Image Generation ~7/~9, Slide Generation ~45/~142, Deep Research ~33/~144, Agent Run ~7/~114, Advanced Agent Runs ~150/~450 — while `glean.com/pricing` 301-redirects to the homepage and no dollar value for a FlexCredit appears anywhere. It also discloses surface-dependent pricing: a Basic Search Query is 0 credits in-product but 1 FlexCredit via the Client API. Consumption is fully forecastable; cost is not. |
| Jasper | Jul 2026 | Introduced credits on top of seats and published a standalone Rate Card (stamped "Updated 6.9.2026") itemising 10 credits per Grid row, 10 per GEO Hub query or page run, 40 per Research or Translation Agent run, 100 per Optimization Agent run, and 1–40 per API/MCP call — with no dollar per credit published anywhere. The meter exists only on the Business plan, whose seat price is also quote-only, so two of the three variables in a Business bill are sales-gated while the buyer forecasts the third. A footnote makes it worse: Advanced Agent runs bill on top of the 10-credit Grid row cost "in both live Grid runs and Grid test mode," so iteration is billable. |
| Photoroom | Jul 2026 | Published absolute monthly credit pools where relative multipliers used to sit — 8,000 / 25,000 / 75,000 AI credits and 1,000 / 3,000 / 10,000 exports for Pro / Max / Ultra — plus per-feature generation ceilings and a Core-tools-vs-Credit-AI-tools split naming exactly what draws from the pool. The Pro/Max/Ultra dollar amounts render as blank fields on the public page and surface only inside app checkout, a state that dates to January 2025. A FAQ adds that "the monthly credit amount included with your plan may vary depending on your country or region," so even the published half is conditional. |
| Outreach | Jul 2026 | The degenerate case at the other end: three Amplify packages (Core 25,000 AI credits, Plus 50,000, Pro 100,000) with platform caps laddering in lockstep (API calls 250k/500k/1M, custom objects 5/10/20, knowledge 500MB/750MB/1,500MB) and "Request pricing" on every tier. Neither half is published — no dollar per credit and no credit per action — only the pool size, which is the one number that tells a buyer nothing on its own. The page still promises "Per user pricing. No platform fees," so the seat, not the pool, remains the commercial anchor. |
| Mintlify | Aug 2026 | Still dollars-only, and the third variable moved instead. Mintlify's $0.01/credit peg is unchanged and it still publishes no per-action consumption rate — the assistant, writing agent, agent skills and automations all draw on one pool with no per-feature figure. What moved is the pool: the free Starter plan's 5,000-credit allotment disappeared from the feature table entirely along with every AI feature, and a reintroduced Pro tier at $450/mo billed annually ($540 monthly) carries 10,000 AI credits at the same $0.01 overage. A buyer who could not compute cost per documentation page before still cannot, and the number that changed is the one neither published half constrains. Reconfirmed on the 2026-08-11 live capture. |
| Linear | Aug 2026 | A vendor publishing BOTH halves, which is the condition this trend says would close the gap. Linear replaced Coding Sessions' one-sided task-tier estimates ($0.50-$1 for a copy tweak, $3-$5 for a small bug fix, $5+ for complex work) with a fully specified equation: model tokens billed at provider-published rates with no markup, plus $0.25 per 20-minute sandbox block. Both the dollars-per-unit and the units-per-action are now derivable, because the units are someone else's published tokens. |
| Recraft | Aug 2026 | The opposite move, and a new failure mode: withdrawing the units-per-period half entirely. Recraft deleted the "30 credits per day" Free-plan figure from both its pricing page and its docs, leaving "Limited free credits" and non-numeric caps ("up to three images per day", "up to two images at a time"). Its paid per-image API unit costs stay published (35 to 120 units per image), so the same vendor publishes a burn rate for paid use and none for free use. |
| Vellum | Aug 2026 | A third variable appears alongside the two halves: EXPIRY. Vellum's docs newly disclose that included usage does not roll over between billing cycles and that purchased pay-as-you-go credits expire 12 months after purchase — terms that change the effective dollars-per-unit without appearing in either published half. Its Custom-plan bundle menu also narrowed from five free-form amounts to three fixed tiers ($25/$45/$115). |
Counterexamples
- ZenRows · Aug 2026 — Seventh both-halves company, and the first to publish both from a standing start. ZenRows replaced dollar-denominated CPM (Scraper API), per-GB (Residential Proxies) and per-session-hour ($0.09, Scraping Browser) pricing with one credits currency across four primitives — and printed the burn table with it: 1 credit for a standard request, 5 for JavaScript rendering, 10 for Premium Proxies, 25 for both. Plan credit totals are published on every rung (Free $0 with 5,000 credits, Build $19-$39, Launch $69-$129, Growth $199-$399, Scale $549-$999 per month across three rungs each, Enterprise above 12.5M credits/month), so dollars-per-credit is derivable exactly as it is at Luma AI. A vendor abandoning three legible dollar units for a synthetic one and publishing the full equation anyway shows the two decisions are independent.
- Diffbot · Aug 2026 — Eighth both-halves company, plus a disclosure of the third variable. Diffbot publishes the consumption side (1 credit per Web Search query, the same as a page extraction) against unchanged plan prices and credit allotments ($0 Free, $299 Startup, $899 Plus, custom Enterprise) with per-credit overage rates, and shipped an interactive workload calculator with sliders for pages extracted, KG entities queried, NLP requests and web searches — the first cost estimator in this file's evidence set. The redesign also surfaced the pool mechanic neither half normally covers: the Free plan does not auto-bill overage at all, returning HTTP 429 Quota Exceeded past 10,000 credits/month or the rate limit until reset, whereas paid plans bill overage automatically pro rata. Complete equation plus a stated overage behaviour is the most buyer-usable disclosure in this file.
- Glean · Aug 2026 — The flagship units-only case started publishing dollars — on an adjacent card. Glean's Model Hub Usage table at docs.glean.com has carried dollar rates since July 2026 and moved twice in this window: on 2026-08-04 it cut GPT 5.6 Terra 20% (input $2.50 to $2.00, cache write $3.125 to $2.50, cache read $0.25 to $0.20, output $15.00 to $12.00) and GPT 5.6 Luna 80% ($1.00 to $0.20, $1.25 to $0.25, $0.10 to $0.02, $6.00 to $1.20), the first published DECREASES on that card; on 2026-08-11 it split GPT Realtime 1.5/2/2.1 into per-modality rows (Audio $32.00 in / $64.00 out, Text $4.00 in / $16.00-$24.00 out, Image $5.00 in) and added GPT-4o Mini TTS ($0.60 text in / $12.00 audio out) and Deepgram Nova-3 Multilingual ($0.0117/min). The same 2026-08-04 release moved Luna from Premium to Standard in the Enterprise Flex tier table so it now falls inside the included 100/user/week allowance rather than always drawing FlexCredits. No dollar value for a FlexCredit is published, so Glean's evidence row still stands as written — but "zero dollar figures" is no longer a true description of the company, and the one-sided cohort is leakier than this trend claimed.
- Snowflake Cortex · Jul 2026 — Publishes both halves at hyperscaler scale, which proves the two-part equation is publishable. A hard dollar anchor — AI Credits at $2.00 for global cross-region routing and $2.20 for regional (data-residency) routing, edition-independent, alongside Platform Credits at $2 / $3 / $4 by edition — plus a per-function Consumption Table covering AI Functions, Cortex Search, Cortex Agents, Snowflake Intelligence, the Cortex REST API and AI Parse Doc.
- Databricks (Mosaic AI) · Jun 2026 — Same completeness: $0.07/DBU for the Artificial Intelligence product line and $0.65/DBU for Model Training, on a published DBU consumption table — Llama 4 Maverick at 7.143 DBU input / 21.429 output per 1M tokens, Llama 3.1 8B at 2.143 / 6.429, GPT OSS 20B at 1.000 / 4.286, BGE Large embeddings at 1.429. Both halves on one page means the ~$0.50 / $1.50 per 1M token cost is derivable without a sales call.
- HeyGen · Jul 2026 — Publishes both halves, and shows why the pair matters. Business credits are $0.05 each in $5 / 100-credit blocks, and the credit-usage schedule grew from 12 to 16 rows (Custom Expressive Motion 40 credits/min, Precision translation 10 credits/min, Seedance2 60 credits at 720p / 150 at 1080p for any length, VEO3 90 credits, AI Model Training 60 credits per training). Because both halves are public, HeyGen's Video Agent increase from 20 to 30 credits/min reads immediately as a 50% price rise even though no subscription price moved — the legibility a single-half disclosure destroys. Its own marketing FAQ still quotes 20 credits/min, so two first-party surfaces disagree.
- Apollo · Jul 2026 — A `gated` company that nonetheless publishes a complete unit equation: per-endpoint credit rates (1–9 credits per enriched person, +8 if a mobile phone is returned, 1 per page of search results, 0 for create/update/list/manage) alongside a published top-up ladder from $0.025 down to $0.015 per credit. Gated seat pricing did not prevent full disclosure of the unit — so the split in this trend is not a proxy for the transparency field.
- Luma AI · Jul 2026 — Posted absolute per-plan credit totals (Plus 10,000, Pro 40,000, Ultra 150,000) beside unchanged $30 / $90 / $300 plan prices and an expanded per-model credit grid, making the dollar side derivable at ~$0.003/credit. Worth noting the redundancy: the 4x / 15x "usage" multipliers on the same cards are just the credit totals restated against Plus (40,000 is exactly 4x of 10,000), so one of the two published figures adds nothing.
- Poe · Jul 2026 — Publishes both halves and still repriced. $19.99 buys a stated 660,000 compute points, and every bot's in-app "Info" button shows that model's exact per-message point cost before you send — an unusually complete consumer equation. On 2026-07-28 the Premium allowance was nonetheless cut 34%, from 1,000,000 points to 660,000, with the price held; the free tier's daily allowance had already been cut ~90% (from ~3,000/day to ~300/day) on 2026-03-30. The lesson: completing the equation still leaves the pool size unguarded.
Trivia
-
Poe is one of the few vendors in this file that publishes BOTH halves — every bot's in-app "Info" button shows that model's exact per-message point cost before you send, and $19.99 buys a stated 660,000 compute points — and on 2026-07-28 it still cut the Premium allowance 34%, from 1,000,000 points to 660,000, with the sticker price untouched. Publishing the equation does not protect the buyer, because the pool size is a third variable that no rate card covers. Jasper's 2026-07-22 rate card sits at the opposite extreme: a 500-row Grid pass is exactly 5,000 credits and no published number converts that into money.
-
LangChain (2026-07-21) and Glean (2026-07-22) published opposite halves of the identical equation one day apart. LangChain collapsed seven per-service meters into two currencies — 1 LCU = $1.50 and 1 LSU = $1.00 — then priced consumption only in infrastructure terms (0.045 LCU/vCPU-hr runtime compute, 0.177 LSU/vCPU-hr database compute) plus a 5–30 LCU band for a single Engine run: a 6x spread on a service scheduled to fire every 6 hours. Glean shipped a 13-line rate card down to p50/p90 per capability while `glean.com/pricing` continues to 301-redirect to the homepage and no dollar value for a FlexCredit exists on any surface.
-
LlamaIndex (2026-06-10) shows that the unpublished half is the one that moves. It raised the credit from $1.00 to $1.25 per 1,000 and trimmed included credits about 20% (Starter 50K to 40K, Pro 500K to 400K) while holding plan prices at $50 and $500 — a 25% per-unit increase and a ~20% allowance cut with no sticker change, leaving every paid plan's price exactly equal to the face value of its included credits. Qodo (2026-06-30) publishes its missing half as a band rather than a rate: 2,500 credits at $0.012 each is labelled "~18 reviews per month," so cost per review is a range by construction.
For buyers
Ask for the missing half in writing before you sign, and treat a refusal as pricing information. If the vendor publishes dollars-per-credit, ask for a burn table for the five actions you will actually run; if it publishes a burn table, ask for the dollar rate. Where you only get one half, insist on a worked example at your expected volume. And add a third question the two-half framing misses: what is the lifetime of the unit? Vellum disclosed in August 2026 that its included usage does not roll over between cycles and that purchased credits expire twelve months after purchase — terms that change the effective dollars-per-unit without appearing in either published half.
For vendors
Publishing both halves is a genuine differentiator right now precisely because so few do, and it costs less than it appears — Snowflake and Databricks do it at hyperscaler scale, so the technical-complexity argument does not survive contact. If you cannot publish a burn table because consumption genuinely varies, publish a distribution rather than a point estimate: Glean's p50/p90 approach is more useful to a buyer than a single average would be, and Clay's 75th-percentile withhold-then-refund is the same instinct applied to billing. What is not defensible is publishing one half and describing the result as simple pricing.
Outlook — what to watch
The gap is closing, but through architecture rather than through vendors choosing to disclose. Linear published both halves in August 2026 in the strongest available form — model tokens at provider-published rates with no markup, plus $0.25 per 20-minute sandbox block — which makes the cost fully derivable because the units are a third party's public tokens. Expect the both-halves cohort to grow via that zero-markup route. Counter-moves continue: Recraft withdrew the units-per-period half entirely from its free tier. If the both-halves cohort keeps growing, this becomes an early-market artefact rather than a structural device.
Bottom line
Ten vendors publish exactly one half of the credit equation and six of them are still counted as transparent. The synthetic unit is where opacity went after the pricing page stopped being where it lived — and the framing needs a third variable, because the unit's expiry moves the price too.
FAQ
Why can't I work out what a credit costs me?
Because you need two numbers and vendors typically publish one. LangChain prints 1 LCU = $1.50 with no per-action conversion; Glean prints p50 and p90 credit burn across 13 capabilities with no dollar figure. Each publishes exactly what the other withholds, so in neither case can you compute a cost per action.
Are credit-priced vendors less transparent than they look?
Often yes, and the corpus metric misses it. Six of the ten one-half publishers still carry price_transparency: public, because their pricing pages do show numbers. A transparency measure that reads pages rather than equations cannot see this.
Is publishing both halves technically hard?
No. Snowflake Cortex and Databricks publish both at hyperscaler scale, so this is a disclosure choice rather than a technical limit. Where consumption genuinely varies, a distribution works: Glean's p50/p90 burn figures are more useful than a single average, even though Glean withholds the dollar side.
Is there a third number I should ask for?
Yes — the unit's lifetime. Vellum's docs newly disclose that included usage does not roll over between billing cycles and that purchased pay-as-you-go credits expire twelve months after purchase. Expiry changes the effective dollars-per-unit without touching either published half, so the equation is really three-part: dollars-per-unit, units-per-action, and how long the unit lives.