Claude API Token Limits: Handling Errors in Production

When building with the Claude API, you will encounter token limit errors in two distinct forms. The first is a 400 error with a message about exceeding the model context window — you are sending too many tokens in a single request. The second is a 429 error indicating you have exceeded your rate limit — too many tokens per minute across all requests. Each requires a different fix.

Per-request context window limits by model

How to fix 400 context limit errors

A 400 error means your prompt plus conversation history exceeds the model context window. The most reliable fix is compressing what you send before it reaches the API. For text content, Token Limits REST API accepts raw text and returns compressed output — call it before sending to Claude. For code and logs, strip timestamps, blank lines, and repeated paths before including them in prompts.

• ✓Compress file contents and logs via POST https://tokenlimits.app/api/compress before including in prompts
• ✓Truncate conversation history to the last N turns instead of sending all history
• ✓Use rolling window summaries — summarize old turns with Haiku, keep only the summary
• ✓Split large documents into chunks and process sequentially instead of all at once
• ✓Use claude-haiku-4-5 for chunked pre-processing steps where full reasoning is not needed

How to fix 429 rate limit errors

Rate limits are measured in tokens per minute (TPM) and requests per minute (RPM). The limits vary by tier and model — check your Anthropic console for your current tier limits. When you hit a 429, the response includes a Retry-After header. Implement exponential backoff starting at the value in that header.

1. Read the Retry-After header from the 429 response
2. Wait that many seconds before retrying (do not hammer the API)
3. Implement exponential backoff: double the wait on each subsequent 429
4. Add jitter (random offset) to avoid synchronized retries across multiple workers
5. Queue requests and process at a controlled rate if you need sustained high throughput

AWS Bedrock and Claude token limits

AWS Bedrock enforces its own token quotas separate from Anthropic's direct API. The default Bedrock quota for Claude models is lower than the direct API and varies by region. You can request a quota increase through the AWS Service Quotas console. Bedrock returns a ThrottlingException (HTTP 429) when limits are hit — handle it the same way as the direct API 429 with exponential backoff.

Using the Token Limits compression API in your app

Token Limits exposes a REST compression endpoint you can call from any language. POST your raw text to https://tokenlimits.app/api/compress with your license key in the request body. The response is the compressed text. Include this in your prompt pipeline before sending to Claude. The API handles up to 100 requests per minute.