Rate limits

Limits apply per IP, globally, and per key. Headers tell you where you stand.

Per IP

Surface	Sustained	Burst
Global API	1000 req/s	4000
/v1/models and aliases	200 req/s	400
Auth endpoints	2 req/s	30
Media endpoints	100 req/s	200
Redeem	2 req/s	20

Concurrency

2000 concurrent in-flight requests globally.
96 concurrent in-flight requests per IP. Beyond that you get an immediate 429 with retry-after: 2, before auth or database work.

Per API key

Field	Type	Description
rpm_limit	requests/min	Returns `429` when exceeded, using an in-memory minute window.
conc_limit	concurrent	Concurrent in-flight cap. Requests queue, then proceed.
daily_cap_micro	micro-USD	Returns `403` when spend reaches the cap. The owner gets at most one email per period.
monthly_cap_micro	micro-USD	Same behavior as the daily cap, on a monthly window.

Security headers

Every response carries:

server: relay
x-content-type-options: nosniff
x-frame-options: DENY
referrer-policy: no-referrer
permissions-policy: camera=(), microphone=(), geolocation=(), payment=(), usb=()
strict-transport-security: max-age=31536000; includeSubDomains

CORS is permissive (*) because the API is token-authenticated and cookieless.

Abuse controls

Control	Behaviour
Signup allowlist	A server-configured set of consumer providers may create accounts; others are rejected.
One account per human	Canonical email collapses dots, +tags and case, enforced in both the backend and the auth worker.
Signup captcha	Verified with the client IP and a hostname allowlist.
Signup reputation	Disposable blocklist → MX → domain age (RDAP), cached 7 days.
Signup daily budget	One created account per IP and per fingerprint per UTC day.
Sign-in throttle	Failures per IP + email, then 429.
Free claim identity	Account ≥ 14 days, current guild member, fresh JWT, captcha.
Free claim uniqueness	One per fingerprint, one per IP, three per ASN.
Daily token caps	Per account per model, so parallel requests cannot overshoot.
Kill switch	Setting an account to suspended or throttled blocks free models with 403.

All blocks are recorded and the high-signal ones are reported to the operator channel. See Free plan for the claim flow.

Headers

x-ratelimit-limit: 4000
x-ratelimit-remaining: 3999
x-ratelimit-after: 3      # only on 429
retry-after: 3            # only on 429

Values reflect the governor layer that handled the request. The global 1000/s budget is the effective outer ceiling. On 429, use exponential backoff seeded by Retry-After, clamped to 1–60 seconds.

PreviousErrors NextTypeScript types