Quick answer
Anthropic measures Messages API capacity with RPM + input TPM + output TPM. Exact numeric capacity depends on the active model and usage tier, so use the Claude Console or Rate Limits API for your organization.
Verified Aug 22, 2026Official source
Current limits
| Constraint | Current value | Scope | Verified source |
|---|---|---|---|
| Messages API limitersThe API uses a token bucket, and limits can be enforced over shorter intervals than one minute. | RPM + input TPM + output TPM | Usage-tier and model specific | AnthropicAug 22, 2026 |
Why does this limit matter?
Input and output token throughput are separate bottlenecks, so a workload with short prompts and long generations behaves differently from retrieval-heavy traffic.
The token-bucket design replenishes capacity continuously rather than at one fixed reset time.
What should you check?
- Open the Rate limits page in Claude Console.
- Inspect retry-after and anthropic-ratelimit-* response headers.
- For organizations, query the Admin Rate Limits API instead of hardcoding account capacity.
Important caveats
- New or limited-history organizations can begin below the standard published tiers.
- Sharp traffic increases can trigger acceleration limits even when a simple per-minute estimate looks safe.
HyperObserve reports the documented platform constraint. Your application, SDK, gateway, provider, region, or account can impose a lower effective limit.
Related