Teams debugging 429 responses on the Claude API often reason from a single number: "our tier allows N tokens per minute, we're under N, so why are we throttled?" The answer is usually one of two mechanisms that sit alongside the headline organization limit. The first is deliberate: workspace-level sub-limits that an admin configured to fence off one team's traffic. The second is automatic: acceleration limits that push back on sharp usage spikes even when quota remains.
Two layers of configured limits
The Claude API enforces two kinds of limits per organization: monthly spend limits and rate limits — requests per minute (RPM), input tokens per minute (ITPM), and output tokens per minute (OTPM) — applied per model using a token-bucket algorithm. A token bucket replenishes continuously rather than resetting on a fixed interval, so capacity comes back gradually as you consume it.
Below that organization ceiling, admins can set custom spend and rate limits on individual workspaces (though not on the default workspace). This is the standard pattern for keeping a batch-analytics job from starving a customer-facing chatbot: give each its own workspace and cap the batch workspace well under the org limit. The organization-wide limit always applies on top — a workspace limit can only be more restrictive, never less. Configured limits can also be read programmatically through the Rate Limits API, which is useful when a platform team wants dashboards that show each workspace's allocation.
Which limit is actually binding? Read the headers
Every response carries rate-limit telemetry: anthropic-ratelimit-requests-*, anthropic-ratelimit-input-tokens-*, and anthropic-ratelimit-output-tokens-* headers with limit, remaining, and reset variants. The important subtlety is that the anthropic-ratelimit-tokens-* headers report the most restrictive limit currently in effect. If a workspace sub-limit binds before the org limit, the headers reflect the workspace number — so a service that "knows" the org tier from documentation can be confused by headers showing a smaller figure. Trust the headers, not the tier table.
When you do exceed a limit, the 429 response names which limit was exceeded and includes a retry-after header with the number of seconds to wait. The official SDKs honor it automatically when retrying. See the rate-limit headers guide for the full header set.
Acceleration limits: the 429 that arrives under quota
Anthropic's documentation adds a caveat that surprises many teams: sharp spikes in usage can trigger 429 errors from acceleration limits even when you have not exhausted your configured rate limit. In plain terms, the platform expects traffic to ramp gradually. A deployment that goes from near-zero to full-quota throughput in seconds — a batch job kicking off, a cache being cold-started across a fleet, a retry storm after an outage — can be throttled during the ramp, regardless of what the workspace or org quota says.
Practical mitigations follow from the cause: warm up traffic in stages instead of all at once, add jitter to fleet-wide restarts, and make sure retry logic uses exponential backoff so a throttled spike does not immediately re-spike. Because ITPM accounting excludes cache reads on most models, a healthy prompt-cache hit rate also flattens how much of a spike ever reaches the limiter — see how cache reads are excluded from ITPM.
Platform caveat: not everywhere
This machinery is a first-party Claude API and Console feature. Claude Platform on AWS uses the same rate-limit schedule but does not offer spend limits or per-workspace rate-limit configuration — organizations there manage cost through AWS billing controls instead, and rate-limit changes go through Anthropic directly (see the Start-tier article). Bedrock, Vertex AI, and Foundry each run their own quota systems entirely.
Where to go next
For the mechanics of the underlying algorithm, read the token-bucket rate-limiting deep dive, or start from the platform overview if you are still choosing a deployment surface.