The redenomination playbook

I have now read well over a hundred AI pricing pages across two editions of the AI Credit Index, most of them twice, with the archived copy beside the live one. The most common change between the two is not a price. It is the meaning of the unit the price is attached to. Seventeen of the 73 reviewed products in Edition 2 carry the pattern: in most of them the rung held, the sticker held, and what a credit buys you changed. This essay is an opinionated guide to doing that well, because the panel shows both the clean version and the other kind, and the difference is entirely in what gets written down.

I am writing this for the people who set the price, not the people who pay it. If you run a credit-metered product you will redenominate at some point, because your costs are external and they move. The question is whether a buyer can still read your page afterwards.

What redenomination looks like in the wild#

Five shapes, each with a name from the panel.

The rename. Pika’s “video credits” became “credits” between 13 and 21 September 2026, with allowances up 29% to 43% and the credits per clip up 65%. Canva stopped metering Standard tools and restated its allowance in Premium-tool uses, so 2,000 became 200 and the series restarts. Speechify went from hours of voiceover a year, to hours of downloads a month, to credits at one per second, across three pages in a year.

The unification. Apollo folded its credit types into one credit in September 2026 (a phone number costs 8, the dialer 2 a minute). Lovable merged separate Cloud and AI dollar grants into one credit balance “rolling out gradually”. Emergent’s “one combined balance” replaced separate credit types. Kiro unified vibe and spec allocations into credits in September 2025. HubSpot consolidated siloed prepaid packs into one pool in June 2025. Twenty-eight products on the panel still carry more than one credit type; the direction is toward one.

The peg. JetBrains made a credit equal to one US dollar from 18 August 2025. GitHub Copilot re-based premium requests to AI credits at one cent from 1 June 2026. monday.com cut its credit price to a cent and raised the credits per action from one to eight, leaving the cost per action unchanged. Eighteen products now peg the credit to a fixed dollar amount, which makes the credit legible and moves every real price into the burn table.

The rewrite of the burn table. Runway and Luma halved credits per second. Rovo rebuilt its table on 31 August 2026. Make published per-model tokens per credit in place of a flat rate. Activepieces replaced a per-model credit table with billing on the provider’s reported dollar cost. Clay moved token-heavy models to variable pricing. The credit stays; the exchange rate per action changes, in both directions.

The redefinition of the unit itself. Freshdesk redefined an email session as a 72-hour window from the customer’s first email. Gorgias made a resolved interaction also count as a helpdesk ticket. Chatbase cut Hobby’s allowance from 2,000 to 500 credits at the same $40 between December 2025 and April 2026. Cursor replaced its dollar allowances with “a set amount of model usage”. These are the ones a buyer cannot see at all without the archive.

How to move a meter without a mutiny covered the change-management side. This is the disclosure side, which the panel suggests is where most of the damage is done.

The disclosure map#

Every fact about a credit lives somewhere, or nowhere, and the Index records where. For the 75 products in Edition 2:

• The burn mapping, credits per action, is on the pricing page for 44 products, in help docs for 24, in the app only for 2, and published nowhere for 5.
• Of the 208 tiers that sell a top-up, 104 price it on the pricing page, 52 in help, 42 only in the app, and 10 not at all.
• Expiry of included credits is stated as expiring on 170 tiers, rolling over on 46, and is unstated on 76. For purchased credits on tiers that sell them, 73 expire, 56 never expire, 13 roll over, and 66 are unstated.
• What happens at zero is a hard stop on 151 tiers, billed overage on 55, a throttle on 19, and unstated on 67.

Each of those “unstated” and “nowhere” counts is a decision a vendor made, usually by not making one. The playbook below is about making them on purpose.

The playbook#

1. Decide which legible shape you are. There are two. Either the credit is pegged to money, in which case every price lives in the burn table and the table must be public, or the credit is a unit of work, in which case the plan card must say how much work. Pika’s September page does the second properly: it prints the burn table, from which a buyer can do the arithmetic: 27 Wan clips or 7 Seedance clips a month on Starter. GitHub’s June page does the first. The illegible shape is a credit that is neither pegged nor mapped: an allowance of “credits” with the credits-per-action table in the app.

2. Put four things on the card. The price. The allowance, with its cadence. The pack price, stated as a rate as well as a pack (“60 credits per $1”, not just “$25”). And what happens at zero: stops, bills, or throttles. Those four are the minimum a buyer needs to compute a bill at a given usage, and the Index’s premium and transparency findings are built from exactly them. Ninety-four tiers on the panel sell a top-up priced only in help docs or in the app, which means the buyer learns the rate after running out or after a search.

3. Put the burn table in the comparison table, dated. Credits per action, by model and by setting, with the date it took effect. Figma’s table is dated “as of August 25, 2026”, which is the right idea; Figma then withdrew its pack tiers and per-credit rate, which is the wrong one. Make’s help page carries its per-model rates with an update date. The table is where a buyer checks whether the same work costs what it did last quarter, and a table without a date cannot answer that.

4. Put the rules in the FAQ, in sentences. Included credits expire or roll over, and when. Purchased credits expire or do not. What counts as a use (Canva: Premium tools only; Notion: agents, Workers, premium models and usage beyond the allowance). Pika’s FAQ does this in two sentences: “Plan credits refresh each billing period and do not roll over. Credits you buy as top-ups stay in your balance until you spend them.” Seventy-six tiers on the panel do not say which.

5. When you redenominate, restart the series out loud. Say what the old unit was and what the new one is, in one line, with a conversion. Pika’s page data carried both numbers, 700 old credits and 900 new ones, in the same JSON object; the conversion existed in the code and not on the page. Canva’s restatement was visible only by comparing captures. The Index handles this with a formerly field and a dated series-restart note, which is more than most vendors offer their own customers.

6. Keep the plan names, or map them. Pika renamed Standard to Starter and Pro to Creator at unchanged prices; OpenArt renamed Advanced to Plus and raised it; v0 renamed Team to Plus; Make went from one plan to three. A buyer with a renewal file is comparing by plan name. If the name changes, say what it was.

7. Date it, and announce it ahead. The clean examples are the incumbents, which is not a coincidence, because their buyers are procurement teams. Notion named the day Workers would start drawing credits and then moved it publicly. Atlassian dated its 31 August table and announced default-on billing for 3 December. HubSpot dated the 16 September pay-as-you-go default. JetBrains blogged the dollar peg on 14 August for an 18 August start. The other kind is a silent redesign: Pika’s landed somewhere in an eight-day window, Chatbase’s allowance cut can only be bracketed across four months of unreadable captures, and Activepieces’ pricing page still showed the old per-model table two weeks after the release that replaced it.

8. Do the arithmetic for the buyer. This is the one most vendors skip and the one that converts. Pika prints the credits a clip costs on each model. monday prints “a typical 3-person team thrives on ~800–1,200 credits per month”. GitHub prints the pool per seat. Kiro prints the credits per plan and the multiplier per model. If a buyer can read the page and compute next month’s bill at their own usage, the redenomination is honest whatever the direction of the change. If they cannot, the Index will do the arithmetic for them, and so will their competitors’ pages.

Why bother#

Because the alternative is being caught. The Index logs burn-rate changes as dated events with the archived page, and the change log now holds 120 dated events. Every silent redenomination on the panel was found, dated to a window, and published, usually with the word “silent” nearby. The vendors that announced theirs got a line in the log that reads like a changelog entry. The vendors that did not got a line that reads like a finding.

A redenomination done with the four card items, a dated table, the rules in the FAQ and a one-line conversion is a price change a buyer can evaluate. Done without them it is how AI hides its prices, and the hiding is now the thing that gets written about.

Every count above is computed from Edition 2’s published observations; the field definitions are in the methodology and the per-product disclosures are on each product’s page.