The meter cut both ways

In June, the first edition of the AI Credit Index read every dated pricing move it had logged and reached one conclusion: the meter only ever tightens. Vendors added meters, hardened caps, raised the credits an action costs, and switched on enforcement. Almost nothing loosened. Four months later the second edition has logged 45 more moves, and the sentence no longer holds. This quarter the meter moved in both directions, and which direction depended on what the vendor was competing on.

That is the finding. The rest of this essay is what it looks like product by product, why it happened, and what a buyer should do about a meter that now cuts as well as squeezes.

Three cuts, in the burn table#

The cuts did not arrive as price cuts. The dollar price of a credit on each of these products is what it was in June. What changed is the burn table: how many credits a given piece of work consumes.

Runway halved the credit cost of its flagship video model. In June, a second of Gen-4.5 cost 25 credits; by 20 September 2026 it cost 12, so the 625 credits in the entry plan bought 52 seconds of video instead of 25. The plan price did not move.

Luma did the same to Seedance 2.0 in Dream Machine. Between 29 June and 9 October 2026 a second of text-to-video or image-to-video went from 47, 107 and 240 credits at 480p, 720p and 1080p to 24, 51 and 126. Verified against both archived pages: a halving at every resolution, at unchanged plan prices.

v0 cut its token rates. The Mini model went from $1 and $5 per million input and output tokens to $0.20 and $1.20; Pro went from $3 and $15 to $2 and $10. The plan prices held; a dollar of included credit now buys between one and a half and five times as many tokens, depending on the model.

These are real price cuts, of a third to four-fifths on the work being bought, and no price tracker will record any of them. The sticker is the same. The credit is the same. The thing that changed is the exchange rate between the credit and the work, which lives in a help page or a burn table and is exactly the channel I described in June as the way vendors hide increases. It hides cuts just as well.

And more variance, in the same channel#

The second group of moves also left the sticker alone, and also changed what work costs. But these did not cut; they made the cost of a task depend on the task.

Clay moved its token-intensive models to variable pricing: “Token intensive models (e.g., GPT-5.1 and Claude 4.6 Sonnet) now have variable pricing based on actual token consumption … no markup.” A long prompt now costs more credits than a short one on the same model.

Activepieces went further. Release 0.92.0, on 24 September 2026, stopped charging AI steps a fixed number of credits per model (the old table: fast model 2, smart 10, frontier 20) and began billing each managed AI call on the dollar cost the provider reports, converted at $0.0005 of model spend per credit by default. The pricing page still showed the old fixed table on 9 October.

Make replaced its flat tokens-per-credit rate with a per-model table: gpt-5-nano at 18,080 input tokens per credit, gpt-5-mini at 3,616, in place of June’s flat 5,000. Cheap models got cheaper per credit and expensive ones dearer, which is the point.

Rovo rebuilt its credit table from 31 August 2026: basic actions at 10 credits per event, premium actions variable, Teamwork Graph and MCP calls at 1 to 10 credits, with the fixed 100-credit Deep Research line gone. And Notion walked a free feature into the meter on a pre-announced date: Workers was “free to try” until credits started counting, first scheduled for 11 August and then moved to 15 October.

None of these is a price rise in the simple sense. Each is a vendor making its meter track its own cost more closely, so that the customer rather than the vendor absorbs the difference between a cheap call and an expensive one. The bill becomes more accurate and less predictable at the same time.

Why the direction split#

Put the two lists side by side and the pattern is not random.

The vendors that cut are competing on how much video or how many tokens a dollar buys, in markets where model prices have been falling for two years. Seedance is not Luma’s model; Runway’s own Gen-4.5 competes against models like it; v0 sells tokens against the falling prices of every other way to buy them. When the input cost falls and the competitor can pass it through, you pass it through, and the burn table is where you do it, because the sticker is the thing you least want to touch.

The vendors that added variance are running their own system on top of models, and their cost per task genuinely varies with the task. A Clay enrichment that calls a frontier model on a long record costs Clay more than one that does not. For eighteen months the answer was a flat credit charge that averaged over the variance. This quarter several of them stopped averaging. That is the COGS argument from AI pricing is just COGS finally showing up reaching the meter: the more the vendor’s cost is external and variable, the more the meter is built to pass it on.

Both groups are telling the truth about their costs. One group’s truth is that the work got cheaper; the other’s is that the work costs what it costs, and that varies.

What the headline numbers can and cannot see#

The Index’s whole-panel median bundled rate fell this edition, from about 1.5 cents a credit to about 1.2. That is not the cuts above, and it is not mainly the 24 new products either: their plans sit almost exactly at the panel median.

So the Edition 2 page also publishes a like-for-like figure, recomputed over only the products observed in both editions. That median fell by the same amount, nearly a quarter. But look underneath it and 35 of those products show no change at all in their median plan rate. The like-for-like fall comes from a handful of re-ladderings: Make returning to three tiers at $10.59, $18.82 and $34.12 for 10,000 credits where June had one plan at $9 for 5,000; GitHub Copilot publishing per-seat credit pools; Pika raising allowances at unchanged prices. The distribution’s middle shifted because the plans on it changed, not because most vendors cut.

And here is the limit of the number: Runway, Luma and v0 do not appear in that like-for-like fall at all. Their dollars-per-credit did not move. The 50% cuts live in the burn table, which no dollars-per-credit statistic can see, which is why the Index logs them as dated events rather than trying to fold them into a rate. The price didn’t change. Your bill did. made that argument about increases. This quarter it applies to the cuts.

What this means if you buy#

Three things, in order of how much money they involve.

Re-read the burn table at renewal, not just the plan. If you buy video generation, the clip you made in June costs half the credits it did on Runway and Luma. Your plan did not get cheaper; your consumption did. That is either a smaller plan next year or more output at the same spend, and a vendor will not volunteer which.

Treat variable-burn products as variable cost. Clay, Activepieces and Make now bill closer to the task. Budget them the way you budget a cloud account, with a monthly check against a fixed unit of work, rather than as a subscription with a known ceiling. The Index flags which products vary burn by model on each product page.

Expect both directions to continue. Model prices keep falling, so the resellers will keep passing cuts through the burn table. Vendors running their own systems will keep making the meter more precise. Neither will touch the sticker if they can avoid it. The only number that stays comparable quarter to quarter is the credit cost of a fixed piece of work, and that is the number worth writing down.

Every move named here is in the change log with its date and its archived page, and the like-for-like figures with their denominators are on the Edition 2 page.