Models

Model ids are namespaced provider/model, and a slug is a contract: it is never repointed at a different model, only at a newer snapshot of the same one. Pinning anthropic/claude-opus-5 in your code is safe. Most models also answer to a bare alias, so claude-opus-5 reaches the same entry.

The table below is generated at build time from the same catalog the gateway routes on, so it cannot name a model Faucet does not carry.

Two answers this page cannot give

Both are ways a page compiled ahead of time could mislead you, so both are worth reading before the table.

A model answers on one endpoint

Faucet serves four kinds of model, and each answers on its own path. Sending a model to the wrong one is a 400 that names the endpoint you actually wanted — not a 404, because the model exists and the slug is not what needs fixing.

Anthropic chat models additionally answer in their native format on /v1/messages. That subset is exactly the chat models whose provider is Anthropic; sending any other model there is a 400.

Reading the capability labels

Each row carries short labels. They mean the same thing here as in the console, so a customer reading both sees one vocabulary.

For a machine-readable version of the same facts, scoped to one model and derived from the functions the routes enforce, read GET /v1/models/{id} instead of this table.

On a chat model:

A chat model with none of these shows text.

On an embedding model, the labels are embeddings and the vector widths it will truncate to, such as 2048/1024/512d. The width is a storage decision rather than a quality one on the models trained for it.

On an image model, the labels name whichever lever decides the price — quality tiers on the OpenAI models, resolution tiers on the Gemini ones, and any size where width and height are free. Each tier consumes a fixed number of output tokens, so the tier you ask for is what you pay.

On a transcription model, the labels are transcription, audio, and the response formats Faucet actually serves for it. That list is narrower than what the provider can produce, and deliberately: a format that carries no usage cannot be metered, so it is refused rather than billed as free. See Transcription.

Reading the rate columns

Each cell carries its own unit, because the models do not all bill in the same one:

• / Mtok — per million tokens. Chat, embedding and the OpenAI and Google image models.
• / min — per minute of audio. The duration-priced transcription models.
• / credit — per provider credit, worth $0.01. The FLUX.2 image models, which report no token counts at all.

An Output column showing — is a column that does not apply to that row rather than a free one: an embedding model returns a vector and has no output tokens, and the per-minute and per-credit models have no output rate at all. A Context of — means the model publishes no token window, which is true of the duration-priced transcription models.

What Faucet deliberately does not carry

Some provider products do not fit a Faucet endpoint or cannot yet be billed honestly, so they stay out of the catalog rather than appearing as broken models:

• Fireworks image generation is asynchronous and has no published per-image price to meter against.
• Voyage rerankers and contextualized embeddings use different endpoints with no OpenAI-compatible shape in Faucet.
• OpenAI audio translation is a separate endpoint Faucet does not expose, and ElevenLabs Scribe Realtime is a WebSocket API while Faucet's transcription surface is HTTP multipart.