What Are AI Credits and How Do They Work?

An AI credit is a unit of account invented by one vendor for one product. It is not a minute, not a word, and not a dollar. It only means something inside the tool that issued it.

That is why credits confuse buyers. Two tools can both sell you "500 credits a month" and give you wildly different amounts of work.

This guide covers what a credit is, how vendors calculate one, whether credits roll over, how to check your balance, what happens when you run out, and how to price a credit plan before you sign up.

What is an AI credit?

An AI credit is a private unit of account that a vendor uses to meter how much of its AI features you may use. You buy credits as part of a plan, and each AI action you run subtracts some of them.

There is no industry standard behind the word. No committee defines a credit, so each vendor sets its own value, its own cost per action, and its own expiry rules.

Treat the word the way you would treat arcade tokens. The tokens are real money once you buy them, but they only spend inside that one arcade.

• A credit is a vendor-specific allowance, not a unit of time, text, or currency
• Credits are usually bundled into a monthly plan, sometimes sold as top-up packs
• Each AI feature has its own credit cost, set by the vendor
• The same word means a different amount of work at every vendor

Why every vendor defines a credit differently

Vendors use credits because their own costs vary by feature, and a credit hides that variation behind one number. A quick text edit and a generated video cost the vendor very different amounts, but both can be priced in the same currency on your invoice.

Credits also let a vendor change prices without changing your plan. Raising the credit cost of a feature is quieter than raising the monthly fee.

The result is a word with no fixed meaning. Design tools, coding assistants, and video generators all sell credits, and none of the numbers are comparable across them.

In design software, credits pay for image and generation actions. In coding assistants, a similar allowance meters requests to the more expensive models. In video tools, credits track seconds of generated footage and the model quality you chose.

• Credits smooth over the vendor's own cost differences between cheap and expensive features
• They let vendors reprice a feature without repricing your plan
• Cheap actions may cost nothing; heavy generation can cost dozens of credits
• Credit values and per-action costs change, sometimes with little notice

Are AI credits the same as tokens?

No. A token is a unit of text the model itself processes, while a credit is a billing unit the vendor invented on top of that.

Tokens are measurable and roughly comparable across model providers. Credits are not, because a vendor can decide that one action costs one credit today and three tomorrow.

Most credit systems are a wrapper. The vendor pays a model provider per token, then charges you in credits with its own margin folded in.

For the mechanics of tokens themselves, read our guide to what AI tokens are.

• Token = a chunk of text the model reads or writes; a real technical unit
• Credit = a billing unit the vendor defines; not technical at all
• You are billed in tokens on a raw API, and in credits inside a packaged app
• A vendor can change a credit's meaning; it cannot change what a token is

Do AI credits roll over?

Usually not. Most vendors reset your credit balance at the start of each billing period, and unused credits are gone.

The published examples are consistent. GitHub Copilot states that unused premium requests do not carry over and that counters reset on the first of each month. Google's help pages say the same about the AI credits attached to its consumer AI plans, and Figma states plainly that AI credits expire monthly and do not roll over.

Top-up packs are the common exception. Credits you buy separately often last longer than the monthly allowance, but the terms differ by vendor and change over time.

People ask this most often about video tools, where credits are expensive and go fast. Kling is a frequent example, and its subscription credits are tied to the billing month rather than banked forever. Check the vendor's own credit policy page before you assume anything, because these rules get revised.

• Monthly plan credits: assume they expire at the reset date
• Purchased top-up credits: often last longer, but read the expiry terms
• Annual plans usually still meter you monthly, not once a year
• Cancelling a plan normally ends access to the credits it included

How AI credits are calculated

Credits are deducted per action, and the size of the deduction depends on which feature you used and which model ran it. Heavier models and longer outputs cost more.

GitHub Copilot publishes this openly: each model carries a multiplier, and your allowance is deducted according to it. Some features cost many requests at once rather than one.

Design and media tools work the same way. Figma's help pages describe cheap actions costing a handful of credits and complex generation costing far more, with image generation varying by the model you pick.

Google's documentation adds a fairer detail worth knowing: the credit cost varies by feature and model, and a failed action does not consume your balance.

• Feature type — text edits are cheap, image and video generation are not
• Model choice — premium models carry a higher multiplier than included ones
• Output length or quality — longer or higher-resolution output costs more
• Bundled actions — one click can trigger a multi-credit job under the hood

How do I find out how many AI credits I have?

Every credit-based tool shows a balance somewhere in the account or billing area, and most also show the date your credits refresh. Figma surfaces an AI balance in its menu. Google shows an AI credits activity page. GitHub shows premium request usage in your account settings.

Check the reset date at the same time, not just the number. A balance without a reset date tells you nothing about how fast you are burning through it.

If you manage a team, look for the admin view rather than your own. Seat-level balances hide the fact that one heavy user can drain a shared pool.

• Look under account, billing, or usage — not inside the AI feature itself
• Note the reset date alongside the balance
• Admins usually get a per-seat breakdown; use it to find heavy users
• Check the balance in week one of a new plan, not week four

What happens when you run out of AI credits mid-task?

You either get blocked or downgraded, depending on the vendor. Neither is pleasant halfway through a job.

Figma blocks: its documentation says that with no credits you cannot complete further AI actions until you get more or wait for the reset. GitHub Copilot downgrades instead, letting you keep working on an included model for the rest of the month.

Ask which behaviour applies before you commit a deadline to a tool. A blocking vendor can stop a client deliverable cold on the 28th of the month.

The fixes are always the same three: buy a top-up, raise the plan, or wait for the reset. Only one of those is free.

• Hard block — no further AI actions until top-up or reset
• Soft downgrade — work continues on a cheaper model, often slower
• Top-up purchase — instant, and usually the worst value per credit
• Plan upgrade — better value if you overrun every month, not once

Per-seat, credits, or per-token API: which billing model is which?

AI gets sold three ways, and credits are only one of them. A per-seat subscription charges by person, credits charge by action, and a raw API charges by token.

Across the tools we buy for client automations, the pattern we keep hitting is that credit plans are easy to start and hard to forecast, while a per-token API bill is intimidating to read and far easier to predict once the work is steady. The credit plans are the ones that surprise us at renewal.

Many businesses end up with all three. That is fine, as long as somebody can say out loud which bill covers which job. For the choice between a seat and direct API access, see our guide to API keys versus subscriptions.

• Per-seat: budget-friendly and boring, which is a compliment
• Credits: convenient for creative bursts, risky for deadlines
• Per-token API: cheapest per unit of work at volume, needs monitoring
• Mixed estates are normal; unlabelled mixed estates are how money leaks

How to work out what a credit plan really buys you

Convert credits into tasks before you pay. The only number that matters is how many real jobs the monthly allowance completes.

Do it in four steps. Name the two or three actions you will actually run, find each one's credit cost on the vendor's pricing or help page, divide the monthly allowance by that cost, then compare the result to your real monthly workload.

If the answer is close, buy the tier above. Credit estimates run optimistic, because retries and rejected outputs still spend the balance.

For per-token pricing rather than credits, our AI model cost calculator and the AI model pricing page do the arithmetic for you.

• Ask the vendor, in writing, whether the plan price includes its credits or bills them on top
• Ask whether unused credits expire, and on what date
• Ask what happens at zero — block or downgrade
• Ask what a top-up costs per credit versus the next plan tier
• Divide the monthly allowance by the credit cost of your most common action