API credits expire, cap you and differ by provider

API credits are money you pay a provider before you use its models, drawn down per token. OpenAI, Anthropic and Google each sell them on different terms.

Those terms decide when you get cut off, whether unused money comes back, and how fast your rate limits grow.

We read the three providers' billing and rate-limit pages on September 15, 2026 and put the rules side by side. Claude Pro and ChatGPT Plus are a different product and never touch this balance.

Credits pay for tokens before you use them

Two billing models exist. With prepaid credits you buy a balance up front and each request deducts its token cost in near real time. With postpaid billing the provider meters usage and charges the card or invoice afterwards, usually monthly. Prepaid is the default for new accounts at all three providers, and Google is the only one of the three that lets a project choose postpay.

Credits are not rate limits. Rate limits cap requests and tokens per minute, and usage tiers cap dollars per month. A large balance does nothing for a low tier, and a high tier does nothing for an empty balance. Both are set out below for each provider.

What a credit buys is the provider's list price. At the rates in our registry as of September 25, 2026, GPT-5.6 Terra is $2/$12 per million input and output tokens, Claude Sonnet 5 is $2/$10, and Gemini 3.5 Flash is $1.5/$9. Our pricing calculator turns those into a monthly figure for a workload you describe.

OpenAI's credits cost $5 at minimum and expire after a year

OpenAI's prepaid billing help page states the minimum purchase is $5 and the default amount is $10. The maximum depends on the account's trust tier. Purchased credits expire after one year, and OpenAI says it cannot extend the expiry of purchased credits or credit grants.

Auto-recharge is on by default when you set up billing. It adds credits when the balance falls below a threshold you set, with a $5 minimum recharge and an optional monthly cap. Turn the cap on before you rely on the feature. Without it, a runaway job recharges as often as it drains.

Credits are treated as non-refundable, with three stated exceptions: a verified billing error, confirmed unauthorized activity, or a confirmed OpenAI service failure or incorrect account restriction.

When the balance reaches zero, requests return a billing error with the code credit_balance_exhausted. OpenAI notes that access may not stop immediately, so a small overrun past zero is possible.

Payments also move you up usage tiers, which cap what you can spend in a month.

Table 1 Download CSV

OpenAI's rate-limits guide, read September 15, 2026. Tiers advance automatically as cumulative payments cross each threshold, and each step usually raises per-model rate limits as well.

Tier	Qualification	Monthly usage limit
Free	Allowed geography, nothing paid	$100
Tier 1	$5 paid	$100
Tier 2	$50 paid	$500
Tier 3	$100 paid	$1,000
Tier 4	$250 paid	$5,000
Tier 5	$1,000 paid	$200,000

Per-model limits at Tier 1 are on OpenAI's free-tier page.

Anthropic's credits are prepaid, and a tier caps the month

Anthropic's support article on prepaid credits states that credits expire one year from the purchase date and that the expiry cannot be extended. All credit purchases are non-refundable. The balance covers API access, the Console playground and Claude Code. You are billed only for successful API calls and completed tasks, so failed requests cost nothing, though a request that timed out while progressing is still charged.

Auto-reload works like OpenAI's: you set the minimum balance that triggers a purchase and the amount to reload to. When credits run out, Anthropic says you can no longer call the API or use the playground until you add more.

New accounts receive a small amount of free credit to test the API. Anthropic does not publish the figure, and larger trials go through sales. Details are on Anthropic's free-tier page.

Spend is capped per calendar month by usage tier, independently of the balance.

Table 2 Download CSV

Claude API rate-limits page, read September 15, 2026. Tiers advance automatically on usage history and account standing, and a higher cap can be requested from the Console.

Tier	Monthly spend cap
Evaluation	Below Start while account history builds
Start	$500
Build	$1,000
Scale	$200,000
Custom	No cap, arranged with the account team

Two behaviours are worth writing into your error handling. Reaching the tier's cap pauses usage until 00:00 UTC on the first of the next month and returns HTTP 429 with the error code enforced_spend_limit_reached and no retry-after header, so automatic retries fail until the month turns. A spend limit you set yourself below the cap returns HTTP 400 instead. And for rate limits, cached input tokens do not count toward input tokens per minute on current models, which makes prompt caching a way to raise effective throughput without a tier change.

Google is the only one of the three that offers postpay

Google's Gemini API billing page describes a free tier that needs no billing account and covers certain models up to their free-tier rate limits. Moving to a paid tier means linking a Cloud Billing account and prepaying at least $5.

From there a project chooses between two models. Prepay buys credits in advance, and usage is deducted from the balance in near real time. Prepaid credits expire after 12 months and are non-refundable, except when the account is switched to postpay. Postpay accrues costs on the Cloud Billing account and charges automatically at the end of the month, or sooner when spending reaches the tier's cap.

AI Studio itself stays free unless you link a paid API key, after which that key's usage is billed. Google prices by input tokens, output tokens, cached tokens and cached-token storage duration, so a long cached context has a storage line as well as a read rate.

Tiers depend on payment history and on time since the first payment.

Table 3 Download CSV

Gemini API rate-limits and billing pages, read September 15, 2026. Rolling ten-minute spend limits also apply: $10 on Tier 1, $50 on Tier 2, $200 on Tier 3.

Tier	Qualification	Billing cap
Free	Active project or free trial	Free-tier rate limits only
Tier 1	Billing account linked	$250
Tier 2	$100 paid, 3 days after the first successful payment	$2,000
Tier 3	$1,000 paid, 30 days after the first successful payment	$20,000 to $100,000 and above

Free-tier model coverage is on Google AI Studio's free-tier page.

One table holds all three rulebooks

Table 4 Download CSV

the provider pages named in each section above, all read September 15, 2026. Where a cell says "not stated," the provider's page carries no figure, and nothing is inferred from third-party posts.

Rule	OpenAI	Anthropic	Google (Gemini API)
Free access	Free usage tier, $100 a month, allowed geographies	Small unpublished signup credit	Free tier, no billing account
Minimum purchase	$5 (default $10)	Not stated	$5 to leave the free tier
Expiry	1 year, not extendable	1 year, not extendable	12 months (prepay)
Refunds	No, with three stated exceptions	No	No, unless switching to postpay
Auto-recharge	On by default, $5 minimum, optional monthly cap	Set a trigger balance and reload amount	Not needed on postpay
Empty balance	credit_balance_exhausted, may not stop instantly	API and playground stop	Prepay deducts to zero, postpay bills monthly
Monthly cap	$100 to $200,000 by tier	$500 to $200,000 by tier, Custom uncapped	$250 to $100,000+ by tier
What raises the cap	Cumulative payments	Usage history and account standing	Cumulative payments plus days since first payment

Credits are not a subscription, and neither is a usage limit

Three products get confused with API credits. A consumer or team subscription such as Claude Pro or ChatGPT Plus is a flat monthly fee for the app, with its own usage limits. It does not add API credit, and API spend does not count against it. Our Claude plans page puts the plan prices beside the API rates and includes a calculator for when an API bill would pass a plan's price, and Claude usage limits covers the app-side caps. GitHub Copilot's "AI credits" are a fourth thing again, a per-plan allowance converted from token usage at Copilot's own rate. Our Copilot pricing page works those through.

Enterprise routes change the picture too. Claude Platform on AWS is billed through AWS Marketplace rather than Anthropic credit purchases, and Anthropic notes that those organizations sit on the Start tier and do not move between usage tiers automatically. A higher limit goes through an account representative.

Five habits keep credits from surprising you

• Put the expiry on a calendar. A year passes, and a balance bought for a pilot is the one that lapses.
• Cap auto-recharge. OpenAI's monthly cap and Anthropic's reload amount are the only brakes on a loop that keeps calling.
• Distinguish the errors. A 429 for a rate limit carries retry-after. Anthropic's spend-cap 429 does not, and OpenAI's exhausted-balance error is a billing error, not a rate limit. Retrying the wrong one burns nothing but time, and alerting on the wrong one hides the real problem.
• Know your tier before launch day. Caps of $100 or $500 a month are hit in an afternoon by a modest production workload. The tier tables above say what you have to have paid, and for how long, to lift them.
• Cache the prefix. On Anthropic, cached reads sit outside the input-token rate limit. On Google, cached tokens carry their own storage price. Both change the bill for one prompt.

A billing rule is a provider decision, and it changes without a version number. Radar tracks pricing and policy changes for the providers you name and keeps a cost scenario with its sources attached, so the next time a cap, a minimum or an expiry moves, you find out from the source and not from a 429.