429 rate_limit_error on Claude Fable 5

Which limit you actually hit

The error message describes which rate limit was exceeded. Read it before you change anything.

• Acceleration limits from a traffic spike

  A sharp increase in your organization's usage can return 429s from acceleration limits even when you are below your published tier numbers. Ramp traffic up gradually and keep usage patterns consistent instead of turning a new workload on at full volume.

• Input tokens per minute

  This is the axis most Fable 5 users hit first, because the pool is 500,000 ITPM on Start against 2,000,000 for Opus 5, and Fable 5's 1M context window invites very large prompts. Only uncached input counts: input_tokens plus cache_creation_input_tokens. Cache reads do not.

• Output tokens per minute

  Fable 5 allows 100,000 OTPM on Start, 300,000 on Build and 800,000 on Scale. OTPM is evaluated in real time against tokens actually produced, so max_tokens does not factor in. Long autonomous Fable 5 runs at high effort consume this axis faster than shorter turns on other models.

• Requests per minute

  Fable 5 allows 1,000 RPM on the Start tier, 2,000 on Build and 4,000 on Scale. The limiter uses a token bucket that refills continuously rather than resetting on a fixed clock, so a burst can exhaust a minute's allowance in a second and trigger a 429 even when your average rate is well under the limit.

• The request went out on the wrong credential

  Limits are enforced at the organization level, with optional lower per-workspace limits. A stray ANTHROPIC_API_KEY in the environment can route traffic through a low-tier key instead of the account you expect, so you hit that key's limit rather than your own.

Fix in this order

Read the response first, then reduce demand, then ask for more capacity. Requesting an increase before you have caching in place usually gets you less than caching would.

1. 1

   3. Cache the repeated part of your prompt

   For most models, cache_read_input_tokens do not count toward ITPM. With a 2,000,000 ITPM limit and an 80 percent cache hit rate you can process 10,000,000 total input tokens per minute. Cache system instructions, tool definitions, large documents and conversation history.

2. 2

   4. Reduce concurrency at the source

   In your own client, cap in-flight requests and queue rather than fan out. In Claude Code, lower CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY, which defaults to 10, avoid running many parallel subagents, and switch to a smaller model with /model for high-volume scripted runs.

3. 3

   2. Identify the axis from the response headers

   Read anthropic-ratelimit-requests-remaining, anthropic-ratelimit-input-tokens-remaining and anthropic-ratelimit-output-tokens-remaining, plus the matching limit and reset headers. Optimizing the wrong axis is the most common wasted effort here.

4. 4

   1. Wait exactly as long as retry-after says

   The retry-after header gives the number of seconds until you can retry. Retries issued earlier than that will fail by definition, so treat the value as a floor rather than a suggestion, and add jitter on top of it across a worker fleet.

5. 5

   5. Ask for a higher limit with real numbers

   Use Request rate limit increase on the Limits page in the Claude Console. Bring your peak input and output tokens per minute per model and roughly what share of your input is cached or repeated context, because that is what the request form is evaluated on.

Related pages