Rate limits

Your API keys and Feather app usage share organization limits for requests, tokens, and concurrency. A key's configured limits apply.

GET /v1/usage returns your organization's limits. Responses report the current allowance through these headers; reset values are durations:

x-ratelimit-limit-requests: 60
x-ratelimit-remaining-requests: 59
x-ratelimit-reset-requests: 12s
x-ratelimit-limit-tokens: 100000
x-ratelimit-remaining-tokens: 90000
x-ratelimit-reset-tokens: 12s

The values above are examples. Allowances refill over time. On 429, wait for retry-after. See errors.

Token reservations

Before generation, we reserve estimated message input plus your output budget. An explicit max_completion_tokens reserves the full maximum for each of n completions; input counts once. If you omit the maximum, we estimate output for rate limiting and use the model's output ceiling for generation.

We adjust the reservation to verified usage, including reasoning. Set max_completion_tokens to your task's output budget. Larger limits reserve more token allowance. Admitted requests consume request allowance even if generation fails.

For higher limits, contact Paper Instruments.