Anthropic API (Claude) Rate Limits

The Anthropic API rate-limits per organization across three metrics: requests per minute (RPM), input tokens per minute (ITPM), and output tokens per minute (OTPM). Standard limits are set by your usage tier (Start, Build, Scale). Exceeding any one returns a 429 with a retry-after header.

Transcribed from the official documentation · verified 2026-07-17

How Anthropic limits work

·Limits are per organization and per model class; different models can be used to their own limits simultaneously.

·Tiers (Start → Build → Scale → Custom) advance automatically with usage; each also carries a monthly spend cap ($500 / $1,000 / $200,000).

·Rate limiting uses a token-bucket algorithm: capacity replenishes continuously rather than resetting on the minute.

·Cache-aware ITPM: for most models, cache reads do NOT count toward input-token limits. With an 80% cache-hit rate, a 2M ITPM limit can carry ~10M real input tokens per minute.

·Sharp traffic ramps can trip separate acceleration limits even below your stated tier limits.

Start tier (per model)

Model	RPM	Input TPM	Output TPM
Claude Fable 5	1,000	500,000	100,000
Claude Opus 4.x (combined)	1,000	2,000,000	400,000
Claude Sonnet 5	1,000	2,000,000	400,000
Claude Sonnet 4.x (combined)	1,000	2,000,000	400,000
Claude Haiku 4.5	1,000	2,000,000	400,000

Opus 4.x and Sonnet 4.x limits are shared across their respective 4.x versions.

Build tier (per model)

Model	RPM	Input TPM	Output TPM
Claude Fable 5	2,000	1,500,000	300,000
Claude Opus 4.x (combined)	5,000	5,000,000	1,000,000
Claude Sonnet 5	5,000	5,000,000	1,000,000
Claude Sonnet 4.x (combined)	5,000	5,000,000	1,000,000
Claude Haiku 4.5	5,000	5,000,000	1,000,000

Scale tier (per model)

Model	RPM	Input TPM	Output TPM
Claude Fable 5	4,000	4,000,000	800,000
Claude Opus 4.x (combined)	10,000	10,000,000	2,000,000
Claude Sonnet 5	10,000	10,000,000	2,000,000
Claude Sonnet 4.x (combined)	10,000	10,000,000	2,000,000
Claude Haiku 4.5	10,000	10,000,000	2,000,000

Custom tier (contact sales) removes the spend cap and negotiates limits individually.

When you hit the limit

·429 error naming which limit was exceeded, plus a retry-after header with the seconds to wait.

·anthropic-ratelimit-* response headers report each limit, remaining budget, and reset time on every request.

·Fast mode (speed: "fast" on Opus 4.8/4.7) has its own separate rate limits and headers.

Getting higher limits

·Console → Settings → Limits → Request rate limit increase (tier advancement is otherwise automatic with usage).

·Prompt caching is the biggest self-serve lever: cached reads don't count toward ITPM on most models.

·Workspaces can be given lower sub-limits to stop one agent from starving the rest of the org.

Anthropic rate limits: common questions

Why am I getting 429s from Claude far below my token limit?

Three usual causes: you hit the REQUESTS-per-minute limit rather than tokens; burst enforcement (60 RPM can be enforced as ~1 request/second, so parallel agent calls trip it); or acceleration limits from ramping traffic too fast. The 429 body says which limit fired.

Do cached tokens count against Claude rate limits?

On most current models, no. Cache reads are excluded from ITPM and billed at ~10% of input price. Only tokens after the last cache breakpoint plus cache writes count. This makes prompt caching the most effective way to raise your real throughput without a limit increase.

Does max_tokens affect Anthropic rate limits?

No. Output limits count actual generated tokens in real time, so there is no rate-limit penalty for setting a high max_tokens.

Hitting these limits from an agent? The specific errors have dedicated fixes: Anthropic / Claude rate limit exceeded (429) · Claude Code “Usage Limit Reached”: Reset Times and Fixes · Claude Weekly Limit Reset Early? Reset Times and Fixes

And if the limits themselves are the problem: this reference is maintained by Standard Compute, a flat-rate gateway with no TPM, daily, or 5-hour caps: a fixed monthly compute budget at full speed until it is used, plus a high anti-abuse request-rate ceiling. How that works →