Model policy & spend caps

Who can use it

Owners and admins of an organization whose plan includes model policy (the Teams plan today). Find it in the dashboard under Model policy.

Model access

• The organization list is either All models the plan allows or a selected set. Each member inherits the organization’s list or gets their own selected set, limited to what the organization allows.
• Matching is exact. Allowing a model also allows its -background variant.
• A blocked model looks like it doesn’t exist: model listings omit it, and a direct call gets the normal unknown-model 404.
• Changes apply within about a minute.

Spend caps (BYOK only)

• Caps apply only to models served on your own provider keys, never to LunaRoute-managed models.
• A cap is per model, set for the whole organization and/or per member, in USD per calendar month (UTC). It resets on the 1st.
• A cap needs a price. Prices come from the synced public price list or from LunaRoute. You can enter your own, and yours wins.
• Before each request, the worst case is reserved: input tokens, plus max_tokens (or the model’s maximum output), at the model’s price. A request that would cross the cap is refused, never truncated. After the response, the actual cost replaces the reservation.
• If a response reports no usage, it counts at the worst case until the nightly reconcile corrects it from recorded usage. A provider error response before any output (for example a 502 from the provider) counts nothing; a stream that fails partway counts its reported usage, or the worst case if it reported none.

When a cap is reached

• 429 with error.code BYOK_SPEND_CAP_REACHED. The body has scope (org or member), model, cap_usd, spent_usd, reserved_usd, request_worst_case_usd and resets_at. There is no Retry-After: the same request will not fit before resets_at or a higher cap, unless a smaller max_tokens brings its worst case under the cap, or in-flight requests (reserved_usd) settle below their reservation.
• 429 MODEL_UNPRICED: the model has a cap but no price. Set one on the Spend tab.
• 503 BYOK_SPEND_GATE_UNAVAILABLE: the spend counter store is unreachable, so capped requests fail closed. Retry later.

See Errors & rate limits for the full list.