“Token limit reached” can describe different problems. A long prompt may exceed the model’s context window; a response may hit its output limit; an API may reject a burst of requests or tokens with HTTP 429; or an account may have reached a billing or usage limit. Read the exact message and, for an API, record the HTTP status and error code before changing anything.
First, identify which limit you reached
- Input or context too large: A request fails when its included conversation, files, instructions, tool data, and requested output do not fit the selected model. A 400 error about context length points here.
- Response stops early: The request succeeds but the answer is incomplete. Check the response’s finish reason, status, and output-token setting. Asking for a shorter answer or continuing in a second request may help; neither raises the model’s maximum output limit.
- Rate limit (often HTTP 429): Too many requests or tokens were sent within a time window. Reduce concurrency and follow the provider’s retry guidance. Rewording a prompt alone does not remove a request-rate limit.
- Credit, spend, or usage limit: Check the account or project’s billing and usage settings. Repeated retries cannot replenish credits or change a configured cap.
For OpenAI API errors, compare the response with the official error-code guide and rate-limit guide. Other providers use their own error names and limits.
If the input is too large
- Remove irrelevant material. Include only the document sections, conversation turns, and tool results needed for the current question. Keep critical instructions and evidence.
- Split large documents by task. Analyze bounded sections, retain source references, then combine checked findings. A summary can lose details, so return to the original source for precise claims and numbers.
- Check the selected model’s current limits. Context windows and output ceilings vary by model and endpoint; use the provider’s live model documentation instead of an old fixed number. A larger context window may help, but it can affect cost and latency.
- Count the actual request. For OpenAI’s Responses API, the input token counting guide describes counting text and multimodal inputs before sending. Simple word or character estimates can miss formatting, images, files, and tools.
In a chat app, starting a new conversation with a short, checked project summary may reduce the context it needs to use. The app decides how much past conversation to include, so do not assume every visible message is always sent to the model. Save important facts outside the chat before starting over.
If the answer is cut off
Ask for a smaller deliverable, such as one section or a short checklist, and request the next section separately. In an API integration, inspect the response metadata and output-token parameter. OpenAI’s conversation-state guide explains how input, output, and, for some models, reasoning tokens relate to the context window. Increasing an output setting cannot exceed the selected model’s own ceiling.
If the API returns HTTP 429
Inspect the error code as well as the status. An OpenAI 429 may reflect request or token rate limits, a rapid traffic increase, exhausted credits, or a spend/usage limit. For a temporary rate limit, pace requests, lower concurrency, and honor Retry-After when present; otherwise use bounded exponential backoff with jitter. For exhausted credits or an enforced cap, resolve the account setting first. Do not loop retries indefinitely: failed requests may still consume rate-limit capacity. The API deployment checklist distinguishes retryable and account-action errors.
Check that the fix worked
Repeat the same task with the reduced input or revised request rate. Confirm the response is complete and accurate, and compare token usage, status, and latency with the failing request. If the message persists, preserve a redacted request ID, model, timestamp, status, and error code for the provider’s support team. Never publish an API key or private prompt in a support post.
