API reference

The main inference endpoint. OpenAI-compatible, with optional models, provider, cache and retry fields an ordinary OpenAI SDK passes straight through. The model id may be a plain vendor/model slug, a slug with a :floor, :nitro or :free routing suffix, multigrid/auto (the cheapest model that fits the prompt and the capabilities you need, with the reason in the response), or @preset/name.

Vectors from any embedding model in the catalogue, billed per token like everything else.

Text-to-image through the same key, balance and spend cap. Billed per image rather than per token.

Billed per unit of output, against the same balance and the same ceiling.

Score a set of documents against a query. Per-unit pricing, not per token.

The catalogue in OpenAI's /v1/models shape, with prices as both a per-token decimal and integer micros. No key required — prices are what people compare before they sign up. ?modality=chat|embedding|image|video|rerank narrows it to the ids one endpoint accepts, which is the list worth trusting over any page.

Every provider that serves one model, with what that route costs and its priority — routes on the same model can be priced differently.

The providers we can route to, and what each one is.

Queue a set of requests to run without holding a connection open. Billed at exactly the same rate as a live call — there is no batch discount. The body is the JSONL file itself, not an uploaded file id: this is not OpenAI's batch API and client.batches.create() will not work against it. ?name= labels the job.

Your jobs, newest first, with counts and status. Capped at 100 with no cursor.

Status and per-item counts for one job.

Stops dispatch. Items that already ran are still billed and their results stay downloadable.

The completed items, including the ones that failed and why, as JSONL keyed by your own custom_id. ?filter=failed|succeeded narrows it. Readable while the job is still running — X-Multigrid-Batch-Status says how far along it was.

What a single request actually did, by request id — for when the response body did not survive a crash or a retry loop. Scoped to your own account.

The same aggregates the Analytics page renders, by day, model and provider, so a finance team can reconcile an invoice without screen-scraping.

Attach your own score to a request id, so Analytics can put quality next to price. Idempotent — a second score replaces the first.

Balance, key cap, account cap and rate limit, so a client can tell which of the four ways it can be refused applies without triggering each one.

The account balance, derived from the ledger.

Requires a provisioning key; an inference key gets a 403 that says so.

Mint a scoped key with its own spend cap and rate limit. The secret is returned once.

Metadata and limits. Never the secret.

Change its name, spend cap or rate limit.

Effective immediately: the key row is read from the database on every request, so there is no cache to wait out.

The tail of the PKCE flow: an authorization code becomes an ordinary inference key with no expiry. Errors here are RFC 6749 shaped, not ours, so any OAuth library parses them. CORS is open — a single-page app is a first-class caller.

The applications this account has registered. Requires a provisioning key, so every client on the consent screen traces back to an account.

Public clients only — PKCE with S256, and no client secret is ever issued or returned.