API reference · ModelBeat docs

ModelBeat is OpenAI-compatible: point an existing SDK at https://api.modelbeat.ai/v1 and change the key. Streaming is supported, and model takes auto or an intelligence tier. See Routing and Limits.

Create a chat completionOpenAI-compatible chat completion. ModelBeat reserves a worst-case cost hold before routing, then settles it against realized usage after the provider responds. A `402` means the hold exceeded the available prepaid balance. Create a text completionOpenAI-compatible text completion. Create a message (Anthropic shape)Anthropic Messages API-compatible endpoint. Accepts Anthropic's wire format; ModelBeat's own catalog-driven routing still decides which model actually serves the request — sending this shape does not pin you to an Anthropic-served model, and does not require one to exist in the catalog. Prompt-cache tokens (`cache_creation_input_tokens`/`cache_read_input_tokens`) are billed at the regular input rate, not Anthropic's cheaper cache-read rate. Create a response (OpenAI Responses shape)OpenAI Responses API-compatible endpoint. Same routing/billing semantics as `/chat/completions` and `/messages` — a different wire format for the same catalog-driven model pool. List the intelligence tiersLists the values `model` accepts, in the OpenAI `/v1/models` list shape, so a stock OpenAI SDK's `models.list()` works unchanged. Authenticated and rate-limited like every `/v1` path, but never metered — no balance hold is taken. During the private beta this is a tier list, not a model catalogue: it returns `modelbeat-advanced`, `modelbeat-standard` and `modelbeat-fast`, and never a model id. Tier names are ModelBeat's own abstraction over the pool, so listing them discloses no model identity (ADR-0070). Create embeddingsOpenAI-compatible embeddings. Real inference — a provider is called and paid for this request — but this path is **not metered by ModelBeat**: no balance hold is taken, no `usage.cost` is attached, and the call does not appear in usage history. Do not build a production integration that depends on this endpoint's cost or usage being tracked. Get the prepaid balanceA standalone, point-in-time balance check — the full breakdown (held, purchased, bonus, signup offer), for a "do I have enough money to fire this batch right now" check before a caller submits a large amount of work. Authenticated and rate-limited like every `/v1` path, but never metered — no balance hold is taken. Get an aggregate spend/usage summaryAccount-level aggregate — total spend, request count, and token count over a date window — merged server-side with a balance snapshot. Authenticated and rate-limited like every `/v1` path, but never metered. Omit both `from` and `to` and the window defaults to the trailing 30 days; the response always echoes back the window actually queried, defaulted or not. For a per-request list instead of an aggregate, see `GET /requests`. List past requestsPaginated, per-request usage history — an OpenAI-style list envelope. Authenticated and rate-limited like every `/v1` path, but never metered. Omit both `from` and `to` and the window defaults to the trailing 30 days, echoed back in the response the same way `GET /usage` does. For a single request by id, see `GET /requests/{id}`. Get a single past requestLook up one request by id. Authenticated and rate-limited like every `/v1` path, but never metered. **`id` must be the `X-Request-Id` response header value from the original request, not the response body's own `id` field.** Same identifier convention as `POST /feedback`. Submit an outcome signal for a prior requestExplicit outcome signal (`"up"` or `"down"`) for a request you made earlier, keyed by that request's id. **`request_id` must be the `X-Request-Id` response header value from the original request, not the response body's own `id` field.** The two are different identifiers, and only the header value is recognized here — the body's `id` returns `404`. Authenticated and rate-limited like every `/v1` path, but never metered: no balance hold, no usage row, no routing.

Everything else the gateway exposes is not part of the API.