Dollar floors go dark, burn rates go public
This trend is weakening. Its founding window sorted cleanly — 10 corpus companies pulled a published price and 14 published one, with dollar-denominated floors coming off the page (Netlify's $500/mo Enterprise floor, Moonshot's $599/yr/seat) and per-action burn rates going on (Glean's 13-capability credit card, Jasper's per-action credits). But on 2026-09-20 Lorikeet published a $75K Signature platform fee in the same change that retired its credit burn table, which meets the trend's own falsification test in letter. Five more dollar floors went dark in September, so the sort is real — it just no longer runs one way.
What's happening — and why
What's happening: ask 'is AI pricing getting more transparent?' as a yes/no question and this window is noise. In the founding window (logged at 380 companies in August 2026), ten companies removed a price they had published; fourteen published one for the first time. That wash is exactly why the sibling trend on the unit rate going dark reached disclosure parity (14 disclosures against 13 withdrawals cumulatively) and weakened this cycle. Sort the same 24 dated events by *what* moved and it stops being noise — it is a trade.
What comes off the page is dollar-denominated. Netlify dropped the 'Custom, starts at $500/month' caption under Enterprise while every other rate held (Free $0, Personal $9, Pro $20, and the credit rates — 15 credits per deploy, 10/GB-hr compute, 20/GB bandwidth, 180 credits per $1 of AI spend). Moonshot pulled Kimi Enterprise's only public price, $599/yr/seat, and routed visitors to sales, while the consumer ladder ($0–$199/mo) and the per-token developer API stayed fully public. Metronome deleted '$100,000 in billing volume included' and '10M events included' three weeks after first publishing them, keeping the flat rates (0.8% of billing volume, $0.04/1k events). Arize removed the AX Pro overage rows — roughly $10 per million trace spans and $3 per GB — while headline prices ($0 Free, $50 Pro) held. Usage AI deleted the worked fee example ($1,200 on $8,000/mo of savings, about 15%). Groq's pricing page now 308-redirects to a homepage with no pricing content at all.
What goes on the page is a burn rate. Apollo replaced a rate-free API pricing page with a per-endpoint credit table (people enrichment 1–9 credits, email waterfall 1–4, phone waterfall 8–25, 0 for create/update/list/manage). Glean expanded its Enterprise Flex docs into a 13-line rate card with p50/p90 credit burn per capability — Deep Research ~33/~144, Advanced Agent Run ~150/~450 — while glean.com/pricing still 301-redirects to the homepage. Jasper layered a credit meter onto per-seat pricing and published the rate card (10 credits per Grid row, 40 per Research Agent run, 100 per Optimization Agent run) with no dollar-per-credit rate anywhere. Luma and Photoroom converted relative '4x usage' multipliers into absolute credit totals at unchanged prices, and Lindy published its credit bands (Plus 3,000, Pro 15,000, Max 35,000 per seat) for the first time.
Why this is worse than either half alone: a buyer needs two numbers to forecast a bill — how many units an action consumes, and what a unit costs. Vendors are increasingly willing to publish the first and increasingly unwilling to publish the second. Glean and Jasper both publish a complete per-action meter with no dollar figure at all: the meter is fully legible and the bill is unknowable. That hands the vendor a private price control while the published half creates an appearance of transparency, because the withheld half is where repricing actually happens. Apollo shows a second-order version — its rate is published, but people enrichment costs 1 credit for demographics or email and 9 if a mobile phone is returned, a 9x swing decided by what Apollo finds rather than what the buyer asked for. HeyGen is the window's one complete equation: it published both the 16-row credit schedule and the dollar rate behind it ($0.05/credit, $5 per 100-credit block).
Why it is weakening: the trend named its own falsifier — a vendor publishing an enterprise floor while withdrawing a burn rate — and in September that happened. Lorikeet (2026-09-20) published a $75K platform fee on its renamed Signature tier, plus FDE Platinum from $25K+, in the same change that removed its credit burn table in favour of dollars per resolution. Klaviyo (2026-09-23) retired Composer's 16,000-credit tiers for a $20/month dollar plan, and Wispr Flow (2026-09-11) replaced a contact-sales Enterprise card with a tier "from $18/user/mo". That followed a late-August warning: Moonshot republished its seat floor at $600/year/seat 24 days after pulling it, and Recraft and Higgsfield withdrew burn rates. The first half of the sort is still strong — five more dollar floors went dark between 2026-09-10 and 2026-09-23 (Deel's $56 Full HR rate, Labelbox's pricing page and calculator, Numeric's $30/user Essentials seat, Digits' $35–$250 per-client ladder, Stability AI's Brand Studio plans) and burn rates kept arriving (Workable's 1 / 2 / 10 credits per evaluate / source / chat; Browserbase's per-plan agent-run counts). But dollar floors re-appeared at the top of two ladders while burn tables were retired at two vendors, so 're-sorting' is real and not one-directional.
How it works
Evidence over time
22 supporting · 4 counter — hover or tap a point for detail, click to jump to the row.
Evidence
| Company | Date | What happened |
|---|---|---|
| Apollo | Jul 2026 | Replaced a rate-free "API Pricing" page (which pushed buyers to a login-gated About Credits page) with a per-endpoint credit table: people enrichment 1-9 credits (1 for demographics/email, +8 if a mobile phone is returned), email waterfall 1-4 (20+ in some configs), phone waterfall 8-25 (45+ in some configs), 0 credits for create/update/list/manage. |
| Glean | Jul 2026 | Publishes no dollar figure anywhere (glean.com/pricing still 301-redirects to the homepage) yet expanded its Enterprise Flex docs to a 13-line rate card with p50/p90 credit burn per capability — Adaptive Reasoning ~11/~38 standard and ~26/~83 premium, Deep Research ~33/~144, Advanced Agent Run ~150/~450. The meter is fully specified; the dollar is not. |
| Jasper | Jul 2026 | Layered a credit meter onto pure per-seat pricing and published a rate card for it — 10 credits per Grid row, 10 per GEO Hub query, 40 per Research or Translation Agent run, 100 per Optimization Agent run, API/MCP 1-40 per call — with NO dollar-per-credit rate published. Seat prices unchanged at $59/$69. |
| HeyGen | Jul 2026 | Published credit top-ups at $0.05/credit ($5 per 100-credit block) and avatar slots at $29/mo (Video Avatar) and $49/mo (LiveAvatar) for the first time, and grew the credit-usage schedule to 16 rows. Video Agent went 20 to 30 credits/min standard, 90 in Seedance mode. |
| Luma AI | Jul 2026 | Replaced relative "4x usage" / "15x usage" framing with absolute monthly credit totals — Plus 10,000, Pro 40,000, Ultra 150,000 — at unchanged prices ($30/$90/$300 monthly). Photoroom made the same multipliers-to-absolutes move on 2026-07-22 (8,000/25,000/75,000 AI credits, 1,000/3,000/10,000 exports). |
| Lindy | Aug 2026 | Published its credit meter for the first time — Plus 3,000, Pro 15,000, Max 35,000 credits per seat with named credit bands by task size — alongside a move to per-seat billing, a 40% Plus cut ($49.99 to $29.99/user/mo) and removal of the struck-through "$8,000/month human assistant" anchor. |
| Netlify | Aug 2026 | Dropped the "Custom, starts at $500/month" caption under the Enterprise tier; the card now reads only "Custom". Every other rate held — Free $0, Personal $9, Pro $20, and the credit rates (deploys 15 credits, compute 10/GB-hr, bandwidth 20/GB, 180 credits per $1 of AI spend). |
| Moonshot AI | Aug 2026 | Pulled Kimi Enterprise's only public price — $599/yr/seat, live through 2026-07-30 — along with its feature list, routing visitors straight to [email protected]. The consumer assistant ladder ($0-$199/mo) and the per-token developer API stayed fully public. |
| Metronome | Aug 2026 | Removed the included-volume numbers from Starter ("$100,000 in billing volume included", "10M events included") three weeks after first publishing them, leaving only the flat metered rates 0.8% of billing volume and $0.04/1k events. The rates themselves held. |
| Arize AI | Jul 2026 | Removed the explicit AX Pro overage rows — ~$10 per million trace spans and ~$3 per GB — during a rebuild into a collapsible feature table. Confirmed absent with every accordion expanded. Headline prices ($0 Free, $50 Pro) and included quotas unchanged. |
| Groq | Aug 2026 | groq.com/pricing now 308-redirects to a homepage with no pricing content and no Pricing nav link. Per-token rates survive in the GroqCloud docs, but three built-in agentic tools (web search, website visits, code execution) have NO live public source at all, because their docs defer to the deleted page. |
| Usage AI | Jul 2026 | Deleted the inline fee calculator (a worked ~15% example: $1,200 fee on $8,000/mo of savings) and the ICR-vs-ESR comparison. The FAQ now says only "a percentage of the savings we realize for you, billed monthly." The only surviving published rate — 20% of EC2 savings, 35% on RDS-class — lives on AWS Marketplace, not usage.ai. |
| Poolside | Jul 2026 | After two years with no pricing page, no free tier and no self-serve path, published its first prices ever — Laguna XS 2.1 at $0.06/$0.12 and Laguna S 2.1 at $0.10/$0.20 per 1M — on OpenRouter and Vercel AI Gateway rather than on poolside.ai, which still 404s /pricing. |
| Observe.AI | Jul 2026 | observe.ai/pricing 404s again, but its AWS Marketplace listing now publishes a fully metered contract with NO seat dimension: Minutes per Month $4.80/unit and Number of Interactions $12.00/unit for 12 months (exactly double for 24 months, i.e. no posted commitment discount). |
| Recraft | Aug 2026 | A burn rate goes DARK, which is the half this trend claims is going public. Recraft removed the specific "30 credits per day" Free-plan figure from both its pricing page and its docs, leaving only "Limited free credits" and non-numeric caps. Its per-image API unit rates stayed published (V4 Styles $0.035/image at 35 units up to V4 Styles Pro Vector $0.12 at 120 units), so the same vendor now publishes a per-action burn rate for paid API use and withholds one for the free tier. |
| Higgsfield | Aug 2026 | Both halves withdrawn at once, which the trend's sorting does not predict. Higgsfield removed the $3 "Special" one-time credit pack — its cheapest entry point and a dollar floor — AND the first-party credits-per-dollar comparison strip (Team $1 = 33 credits, Max = 31, Pro = 27, Basic = 14), a published burn rate. Both had first appeared on 2026-08-16. Only a generic "X% CHEAPER" badge on Max remains, and subscription tier prices are unchanged. |
| Deel | Sep 2026 | The top Deel HR tier lost its $56 per employee list price and became "CUSTOM PRICING" while the mid tier was cut $30 to $22 — a dollar ceiling gone dark. |
| Workable | Sep 2026 | Published a per-action burn table for the Workable Agent (1 / 2 / 10 credits per evaluate / source / chat) — a burn rate going public. |
| Labelbox | Sep 2026 | labelbox.com/pricing, its calculator and /plans all 404; the $0.10/LBU rate survives only in a docs Limits table — the dollar surface gone dark. |
| Numeric | Sep 2026 | Removed its only public price, Essentials at $30/user/month, for a quote-only suite. |
| Digits | Sep 2026 | Pulled its public per-client firm ladder ($35 / $65 / $100 / $250 per client/mo) for partner-only plans with no prices shown. |
| Stability AI | Sep 2026 | Brand Studio's public plans page (Creator $19/mo, Core $50/mo) now 404s; pricing sits behind a login. |
Counterexamples
- Zilliz Cloud (Milvus) · — — Broke the pattern by publishing plain DOLLARS, not burn rates: on 2026-07-21 it posted entry prices for its paid plans for the first time — Standard from $0/mo Serverless or $126/GB/mo Dedicated, Enterprise from $197/mo Dedicated — while adding a per-minute On-demand Compute meter.
- Freshworks · — — Also published a plain dollar rate after two years of sales quotes: Freddy AI Agent at $49 per 100 bot sessions (~$0.49/session), identical on Growth, Pro and Enterprise (2026-07-21) — in the same release where it deleted its free plan.
- You.com · — — Replaced an open-ended ">$2,000.00 /1k calls" label on its top Research tier with a firm $1,200.00 /1k calls on 2026-07-22 — simultaneously a ~40% cut at the top of the range and a disclosure, while keeping the Contact Sales CTA.
- Moonshot AI · Aug 2026 — DIRECT COUNTEREXAMPLE to this trend's central sort, at one of the six companies it cites for withdrawing a floor. Moonshot republished exactly the kind of item this trend says is going dark — a dollar-denominated per-seat floor — at $600/year/seat, minimum 2 seats, with a working self-serve Subscribe checkout, 24 days after pulling $599/yr/seat for a sales-only motion. The floor came back, and it came back with a checkout attached rather than as a "starts from" caption.
- Wispr Flow · Sep 2026 — Replaced its contact-sales Enterprise card with a self-serve Growth/Enterprise tier "from $18/user/mo" (annual) — an enterprise floor published.
- Lorikeet · Sep 2026 — Published a $75K platform fee on its renamed Signature tier (plus FDE Platinum from $25K+) in the same change that removed its credit burn table for dollar rates — the trend's falsification condition met in letter: a floor published while a burn rate was withdrawn.
- Klaviyo · Sep 2026 — Retired Composer's credit burn (16,000-credit tiers) for a $20/month dollar plan and restated free allowances in dollars — burn table out, dollars in.
Trivia
-
Lorikeet (2026-09-20) now prints a $75,000 platform fee on its top tier — one of the few enterprise floors published anywhere in the corpus — in the same change that retired its credit table.
-
Five dollar floors went dark in 14 days (Deel, Labelbox, Numeric, Digits, Stability AI; 2026-09-10 to 2026-09-23) while two burn tables were retired for dollars (Lorikeet, Klaviyo).
-
Glean and Jasper both publish a complete per-action meter with no dollar figure anywhere. Glean's Enterprise Flex docs specify p50/p90 credit burn for 13 capabilities — Deep Research ~33/~144, Advanced Agent Run ~150/~450 — while glean.com/pricing still 301-redirects to the homepage. Jasper published per-action credit costs (10 per Grid row, 100 per Optimization Agent run) with no dollar-per-credit rate at all.
-
Apollo's per-endpoint credit table prices an outcome the caller cannot control: people enrichment costs 1 credit for demographics or email and 9 if a mobile phone is returned — a 9x swing decided by what Apollo finds, not by what the buyer asked for.
-
Poolside published its first prices in two years — Laguna XS 2.1 at $0.06/$0.12 and Laguna S 2.1 at $0.10/$0.20 per 1M — on OpenRouter and Vercel AI Gateway rather than on its own site, where /pricing still 404s. Observe.AI did the same thing: its own pricing page 404s while its AWS Marketplace listing publishes a fully metered contract at $4.80/minute and $12.00/interaction with no seat dimension at all.
For buyers
Treat 'published pricing' as two separate questions and ask both: what does an action consume, and what does a unit cost? A vendor that publishes only the meter has reserved the right to reprice you without editing a page anyone can diff — so get the dollar-per-unit rate into the contract even when the burn table is generous and the tier price is public. Where the burn rate depends on what the vendor finds rather than what you request, contract on the p90, not the headline: Apollo's 1-credit and 9-credit people-enrichment cases are the same API call, and Glean at least publishes both percentiles (Deep Research ~33 p50 vs ~144 p90). Before concluding a price is private, check the cloud marketplace — Observe.AI's own page 404s while its AWS listing carries $4.80/minute and $12.00/interaction, Poolside's only published prices ($0.06/$0.12 and $0.10/$0.20 per 1M) live on OpenRouter and Vercel AI Gateway, and Usage AI's only surviving rate (20% of EC2 savings, 35% RDS-class) is on AWS Marketplace. Finally, capture withdrawn numbers before they vanish: Netlify's $500/mo floor, Arize's ~$10/M-trace-span overage, Metronome's $100,000 included allotment and, since September, Numeric's $30/user seat and Digits' $35–$250 per-client ladder are now negotiation anchors that exist only in archives — and withdrawn floors can come back (Moonshot republished its seat at $600/year 24 days after pulling it).
For vendors
Publishing a meter without a rate buys less trust than it looks like it does — sophisticated buyers read a rate card with no dollar on it as a repricing option, not as transparency. If you run this play, at minimum publish percentiles rather than averages (Glean's p50/p90 per capability is the honest shape) and keep the schedule current as features change, which is real product work: HeyGen grew its credit-usage schedule to 16 rows and had to restate Video Agent from 20 to 30 credits/min, 90 in Seedance mode. Absolute allowances beat multipliers — Luma and Photoroom both replaced '4x usage' framing with hard credit totals at unchanged prices, which costs nothing and removes an entire class of support ticket. The stronger position is to publish both halves: HeyGen's $0.05/credit and Lindy's per-seat bands make the bill computable, and the counter-camp shows plain dollars still win discovery (Zilliz posted $126/GB/mo Dedicated and Enterprise from $197/mo, Freshworks put $49 per 100 bot sessions on a previously quote-only agent, You.com replaced an open-ended '>$2,000/1k calls' with a firm $1,200), and in September Lorikeet and Klaviyo went further, retiring their credit burn tables for dollar pricing outright.
Outlook — what to watch
The founding trade was clean: in the first window no corpus company ran the sort backwards. The fresh-window test it then set has now come back mixed. The withdrawal half kept going — five dollar floors went dark in 14 days (Deel, Labelbox, Numeric, Digits, Stability AI) — and burn rates kept arriving (Workable, Browserbase). But the falsifier fired in letter at Lorikeet, which published a $75K Signature platform fee while retiring its credit burn table on 2026-09-20, and Klaviyo retired Composer's 16,000-credit tiers for a $20/month dollar plan three days later; Wispr Flow published an enterprise floor from $18/user/mo. Add Moonshot's August republication of its seat floor and Recraft's and Higgsfield's burn-rate withdrawals, and both halves have now been observed moving in every direction. What would firm it back up: another window in which withdrawals stay dollar-only and no vendor swaps a burn table for dollars. What would supersede it: more vendors following Lorikeet and Klaviyo, which would make 'meters out, dollars in' a competing trend rather than an exception. Glean and Jasper remain the tell — neither has attached a dollar to its published meter.
Bottom line
Dollar floors going dark while burn rates go public is weakening as a one-way trade. The first half is still strong — five more dollar floors vanished between 2026-09-10 and 2026-09-23 — but Lorikeet published a $75K enterprise floor while retiring its credit burn table on 2026-09-20, meeting the trend's own falsification test in letter, and Klaviyo swapped credits for a $20/month dollar plan. Keep asking for both halves of the price equation; just don't assume the dollar half is the one that will disappear.
FAQ
Is AI pricing getting more transparent or less?
Neither — it is re-sorting. In one 2026 window across a 380-company corpus, 10 vendors removed a price they used to publish and 14 published one for the first time, so the binary count is a wash. What changed is the composition: dollar-denominated floors and ceilings (Netlify's $500/mo Enterprise floor, Moonshot's $599/yr/seat, Arize's ~$10-per-million-trace-span overage, Metronome's $100,000 included allotment) came off, while per-action consumption rates went on.
A vendor publishes a full credit rate card but no dollar price. Is that transparent?
No — it is the weakest of the three states, worse than publishing either half alone. Glean specifies p50/p90 credit burn for 13 capabilities while glean.com/pricing 301-redirects to the homepage; Jasper published per-action credit costs (10 per Grid row, 100 per Optimization Agent run) with no dollar-per-credit rate anywhere. In both cases the meter is fully legible and the bill is unknowable, which means the vendor can reprice you by changing what a credit costs without editing a page. Ask for the dollar value of a credit, in writing, and for whether that ratio is contractually fixed for the term.
The vendor's pricing page 404s. Is the price actually private?
Check the cloud marketplace and the model routers first. Observe.AI's own pricing page 404s while its AWS Marketplace listing publishes a fully metered contract — $4.80 per minute and $12.00 per interaction, with no seat dimension and, notably, no commitment discount (the 24-month price is exactly double the 12-month one). Poolside published its first prices in two years on OpenRouter and Vercel AI Gateway while poolside.ai/pricing still 404s, and Usage AI's only surviving published rate (20% of EC2 savings, 35% on RDS-class) is on AWS Marketplace rather than usage.ai.
Is this trend still holding?
Only partly, which is why it is marked as weakening. Dollar floors kept disappearing in September 2026 — Deel's $56 Full HR rate became Custom, Labelbox's pricing page 404'd, Numeric dropped its $30/user seat, Digits pulled its $35 to $250 per-client ladder and Stability AI put Brand Studio pricing behind a login. But the trend said it would be falsified if a vendor published an enterprise floor while withdrawing a burn rate, and Lorikeet did exactly that on 2026-09-20 with a $75K Signature platform fee. Klaviyo also retired credit burn for a $20/month dollar plan. The trade still happens; it no longer runs only one way.