Rate limits

Three ceilings are enforced today: 1,000 requests per principal per UTC day, 32,768 output tokens per request, 524,288 input characters per request. Paid answers are not counted against the daily number; the per-request ceilings are the same.

Enforced ceilings (/v1/models → freeTier)

{"requestsPerDay": 1000,
 "maxOutputTokens": 32768,
 "maxInputCharacters": 524288}

429 {"error": {"code": "free-quota-exhausted"}}   // resets at UTC midnight, no Retry-After

The assurance ladder (tokens/day)

Each verification rung publishes a daily token allowance (assurance on /v1/research/status, and /v1/secure). **It is not enforced today** — the three ceilings above are what actually apply. The rung decides the band it opens (code-review / cyber / offensive) and the firewall's minimum rung.

rung	tokens/day	ceiling	band
`card` · L1	102400	102400	code-review
`identity` · L2	50000	200000	code-review
`business` · L3	2000000	8000000	cyber
`contracted` · L4	8000000	32000000	offensive

Concurrency and retries

There is no concurrency limit; concurrent requests each count as one.
Resending the same body re-attaches to the same job and is not counted (deterministic job id, within 6 hours).
Requests refused by guardrails, and models whose route is not configured, are not counted.

Past the free allowance: with an AI balance the same request simply runs paid (Per-request cost); without one, the next UTC day.