Four limits can apply to one API call
A free API may advertise one headline number while enforcing several independent limits. Your request must fit all of them. Reaching any one limit can stop the call even when the dashboard still shows remaining credits.
| Limit | Controls | Typical symptom |
|---|---|---|
| Credit balance | Total spend available | Requests stop when the wallet reaches zero |
| Token allowance | Total text processed or generated | Long prompts consume the quota quickly |
| Requests per minute | Burst speed | HTTP 429 during rapid testing |
| Tokens per minute | Throughput | Large prompts throttle even with few requests |
| Context window | Maximum size of one request | Oversized prompt or conversation is rejected |
Convert the offer into a test budget
Do not estimate from the advertised balance alone. Choose the exact model, find its input and output prices, and measure a representative request. Output tokens are often more expensive than input tokens, so a chat response or code generation can cost more than the prompt suggests.
- 1Choose the model and record its current input and output prices.
- 2Run five representative requests and log input tokens, output tokens, latency, and errors.
- 3Calculate the average cost per successful request.
- 4Divide the free balance by that average, then reduce the result by 20 percent for retries and variation.
- 5Confirm whether the allowance resets or disappears permanently after it is used.
One-time credits and recurring allowances solve different jobs
One-time signup credits are useful for evaluation, benchmarking, and a short prototype. They are not recurring infrastructure. A monthly or weekly allowance is more suitable for a small continuing project, but only when the reset and rate limits are documented.
Some gateways provide selected zero-priced model routes instead of a wallet balance. Record the precise model identifier. A similarly named paid route may charge the wallet even when another route is free.
- Use one-time credits for experiments with a clear ending.
- Use recurring allowances for small predictable workloads.
- Use open-source local models when a project must not depend on a promotional balance.
- Never place a secret API key in browser code or a public repository.
Protect yourself from an accidental bill
Use provider spending limits when available, isolate development keys, and monitor usage by project. If a provider automatically switches to paid usage after free credits, set a hard budget or remove the payment method before experimenting.
For public applications, add server-side rate limiting and abuse protection. A free allowance that is generous for one developer can disappear quickly when an exposed endpoint is called by bots or copied into another project.
Common questions
How many requests is one million tokens?
There is no universal answer. It depends on the average prompt, response length, tokenizer, and whether the allowance counts input, output, or both.
Does a free workspace make every model free?
Usually not. A gateway account or workspace can be free while individual model calls retain their normal provider price.
What does OpenAI-compatible mean?
It means the provider follows familiar API request shapes and SDK patterns. It does not mean the service is operated by OpenAI or that every OpenAI feature is supported.