A Credit Is Not a Currency: Two Vendors Rewrote the Exchange Rate on the Same Day

On 25 August 2026 two unrelated companies made the same kind of change on the same day. Make updated how its AI provider converts tokens into credits, moving from a single combined rate to separate rates for input and output. Figma updated its AI credit add-ons so that the same money buys more credits, roughly twice as many on the Professional plan and about 1.6 times as many on Organization and Enterprise. Neither announcement raised a price. Neither changed a subscription tier or a seat cost. Both silently rewrote the exchange rate between work performed and balance consumed, which for anyone forecasting AI tooling spend is the only number that matters. The credit, it turns out, is not a currency. It is a vendor maintained conversion table that can be revised on a Tuesday.

Look at the Make change closely, because the structure of it is more interesting than the direction. Previously credits were calculated on total tokens, input and output combined. From 25 August the two are metered separately, and the rates are not close. On the small and medium tiers one credit buys 18,080 input tokens or 2,260 output tokens. On the large tier one credit buys 3,616 input tokens or 452 output tokens. Make gives the example of a data extraction step that cost about 0.6 credits on the medium tier under the old rates and about 0.16 credits under the new ones, a reduction of roughly 73 percent. That is a genuine saving and it was presented honestly. It is also entirely workload dependent. Automations that read a lot and write a little got dramatically cheaper. Automations that generate long outputs did not move nearly as much, and the gap between the two just widened by a factor of eight on the large tier.

The Figma change points the other way and lands in the same place. Existing add-on subscriptions and pay as you go arrangements were uplifted automatically, the cost stayed flat, and the credit balance went up. Good news, and worth taking. But if your design team bought an add-on in July because the included allowance was running short, the shortfall that justified the purchase may not exist any more, and nobody in finance will be told. Every forecast built on the old ratio is now wrong in the pleasant direction, which is exactly the kind of wrong that never gets investigated. A budget that comes in under is treated as a budget that worked, right up until the next rate revision goes the other way.

The reason these two moves belong in the same article is that they expose a property of AI tool billing that most teams have not internalised. A Figma credit, a Make credit, a GitHub AI Credit and a Cursor request are four different units that share one word. GitHub is the useful exception here, because when Copilot moved to AI Credits on 1 June 2026 it pinned one credit to one US cent, which makes the denominator stable and the arithmetic checkable. Everywhere else the credit floats against tokens, against model tier, and now against whether those tokens went in or came out. You cannot add these units together, you cannot compare two vendors by credit price, and a consumption chart drawn across a rate change is measuring two different things on one axis.

What this breaks in practice is ordinary and expensive. Forecasting off historic credit consumption stops working the moment the rate moves, and the movement is not announced on your invoice. Chargeback across teams stops being fair, because a team doing extraction heavy work in Make just got cheaper relative to a team doing generation heavy work, for reasons that have nothing to do with either team. Procurement comparisons built on credits per dollar go stale without any signal. Worst of all, the basic operational question, did our AI usage go up this month, becomes unanswerable from the billing data alone. The fix is unglamorous: record consumption in the underlying units the vendor exposes, tokens or runs or generations, and treat credits as the billing wrapper rather than the measurement.

There is a governance dimension that is easy to skip because this looks like a finance topic. A change to how a supplier meters and charges for a service is a change to the commercial terms of that service, and it is the kind of change that vendor management controls under SOC 2 and ISO 27001 exist to catch. In practice almost no organisation re-reviews a vendor because a help centre article updated a conversion table. Yet where overage is enabled, the conversion rate is also the rate at which an unattended agent can spend your money, which makes it an availability and cost exposure control rather than a footnote. ISO 42001 asks for an inventory of the AI systems in use with a named owner and defined limits for each, and a metering change is precisely the kind of drift that inventory is supposed to surface. Vanta, Drata and Secureframe can hold that register as evidence against a supplier monitoring control, which is more useful than a spreadsheet nobody outside finance opens.

The work here is small and worth doing this week. List every tool in your stack that meters AI in credits or units rather than flat seats, which for most teams now means Cursor, GitHub Copilot, ChatGPT and Claude at the API and team tiers, Figma, Make, and the app builders like Lovable, v0 and Bolt, plus generation tools such as ElevenLabs and Midjourney. For each one, write down four things: what a unit is defined as today, what the current conversion rate is, what happens when the balance hits zero, and who owns the number month to month. Then subscribe someone to the pricing page or changelog for each vendor, because that is where these changes are published and it is not your inbox. Teams on Copilot should pair this with a look at their base allowance, since the promotional credits that cushioned the June transition have now lapsed and the September cycle is the first full month at the real rate.

None of this is an argument against credits. Metering is the correct model for agent workloads, flat seat pricing was always cross subsidising the heaviest users, and both the Make and the Figma changes made their customers better off. The discipline that is missing is treating a credit as what it is, a vendor defined unit that can be restated without notice, rather than as money. Two companies restated it on the same day in August, quietly, for good reasons, and in opposite directions. Whichever way the next one goes, the person who owns your AI spend should learn about it from your own monitoring rather than from a variance nobody can explain.