AI credits are the unit almost every enterprise vendor invented in 2026 to charge you for AI features. You buy an allotment, your usage draws it down, and when it runs out you either stop or start paying overages.
The pitch is that credits give you predictability. What they mostly give you is a currency you cannot price, cannot compare, and cannot forecast.
Every vendor minted its own money
SAP has AI Units. Oracle has AI Units, which are not SAP’s. Microsoft has Copilot Credits and ACUs. Workday has Flex Credits. ServiceNow meters Assist consumption, AWS has Bedrock AgentCore, Google has Agent Engine.
None of them are comparable. Each meters differently, converts differently, and expires under different rules. A credit is not a unit of anything. It is a private currency issued by the company selling you the goods, at an exchange rate it sets and can change.
That is the part worth sitting with before you evaluate any of it on price. You are not comparing rates. You are comparing seven currencies against a basket of goods nobody has itemised.
What a credit is actually pegged to
Usually tokens, sometimes actions, occasionally both, and the conversion is frequently unpublished.
Tokens are the underlying unit of model usage: text going in, text coming out, plus cached context. Most buyers have no intuition for how many tokens a task consumes, which is the root of the problem. Nobody can estimate an annual spend denominated in a unit they cannot picture.
Vendors are candid about this in places. Abacus AI states plainly that its credits are not tokens and that it does not publish exact per-model credit rates, because the rates keep changing. That is honest, and it also means the price of the thing you are buying is not knowable at the time you buy it.
Where a vendor does publish, read carefully for three things: whether cached and input tokens are billed at the same rate as output, whether credits expire monthly or roll over, and whether the model you actually use is priced differently from the default one in the marketing table.
The bills that made this famous
Zylo’s 2026 SaaS Management Index surveyed 218 IT leaders and found 78% had hit unexpected charges tied to AI or consumption in the past year. Not a fringe experience. The majority.
The GitHub Copilot transition on 1 June 2026 is the case study, because it was well signposted and still went badly for heavy users. GitHub replaced premium request units with AI Credits metered on token usage, and held base plan pricing exactly where it was. Reporting since has described individual bills rising as much as sixtyfold, from $29 to around $750 and from $50 to roughly $3,000. Uber is reported to have exhausted its 2026 AI coding budget by April.
Those individual figures come from secondary coverage rather than from GitHub, so treat the multiple as illustrative. The direction is not in dispute, and the mechanism is the point: the plan price did not change at all. The plan price was never where the money was.
Why credits exist
Not villainy. Inference costs real money per use, inside a business model built on near-zero marginal cost, and a flat subscription over variable cost is a losing trade for the vendor when usage is unbounded.
Credits solve a genuine problem. They also happen to solve several others in the vendor’s favour: revenue that grows without a renegotiation, a price that can be changed by adjusting a conversion rate rather than a published number, and breakage on unused allotments. A mechanism can be both necessary and asymmetric.
What to ask before you sign
- What is the conversion rate, in writing, and can you change it during our term? The single most important question, and the one most likely to get a vague answer. A rate the vendor can move unilaterally is a price they can raise without telling you.
- Do unused credits roll over or expire? Monthly expiry means you are buying peak capacity every month and discarding the difference.
- What happens at zero? Hard stop, automatic overage, or a call from your account manager. Each has a very different failure mode, and the hard stop is not always the worst one.
- Can we cap it? A hard ceiling you control, not an alert. Alerts arrive after the money is spent.
- What does a typical week cost for one heavy user? Make them model it. If they will not, that tells you they cannot, which tells you something about your own forecast.
Instrument it before you need to
Whatever you agree, assume the first two months of actuals will surprise you, and set alerts at 50%, 75% and 90% of the allotment so you find out while there is still time to change behaviour.
Then watch for the change nobody announces. Conversion rates move, allotments get restructured, and a model swap in the background can change your consumption without any decision on your side. That is what CopperFeed records: dated entries for repricings and billing model changes as they land.
Related: How to Budget for Usage-Based Pricing, and The End of Per-Seat Pricing.
Figures here come from published research and reporting, attributed inline. Individual bill increases are drawn from secondary coverage and not confirmed by the vendors involved. Verify current terms before acting on any of it.