Model capabilities

GET /v1/models/{id} returns OpenAI's four model fields plus a faucet block describing what this gateway will serve for that model. That is not the same as what the model can do, and the difference is the reason the block exists: whisper-1 genuinely produces srt, and this gateway refuses it, because a bare document has nowhere to report usage.

Branch on this rather than hard-coding anything from these pages. The block is produced by the same functions the routes enforce, so it cannot drift from behaviour — a capability document that can drift is worse than none, because clients believe it.

curl https://api.intfaucet.com/v1/models/elevenlabs/scribe-v2 \
  -H "Authorization: Bearer $FAUCET_API_KEY"

{
  "id": "elevenlabs/scribe-v2",
  "object": "model",
  "created": 0,
  "owned_by": "elevenlabs",
  "faucet": {
    "kind": "transcription",
    "response_formats": ["json", "faucet_json"],
    "supports_streaming": false,
    "max_file_bytes": 26214400,
    "surcharged_parameters": [
      "entity_detection", "entity_redaction", "keyterms", "detect_speaker_roles"
    ]
  }
}

Two things about the envelope itself:

• The block is on the detail route only. GET /v1/models stays exactly OpenAI-shaped for clients that only enumerate.
• An alias is echoed as the canonical slug. Ask for claude-opus-5 and id comes back anthropic/claude-opus-5, so a client that stores what it reads back stores the stable name.

created is always 0. OpenAI's field is a real publication timestamp Faucet does not track, and a moving value would make every poll look like a brand-new catalog.

kind: "transcription"

response_formats is the field to read before sending response_format. It is the same function the audio route rejects with and the same one the Models reference labels rows from, so the three cannot disagree. faucet_json appears here when the gateway can normalize that model's output — no model produces it, so no catalog entry lists it.

surcharged_parameters names parameters and publishes no rates, which is a decision rather than an omission: the basis points behind them are not yet confirmed against a real invoice, and a number here would be a machine-readable promise about a bill that cannot be backed. See Transcription for what they cost today and the caveats on those figures.

kind: "embedding"

The embeddings route runs the same derivation before it resolves a credential, so asking for a width outside the published set is a local 400 rather than a provider-dependent surprise.

kind: "image"

Chat models publish no block

A chat model is reachable through two endpoints — /v1/chat/completions and, for Anthropic models, /v1/messages — and the passthrough and translated dialects refuse different things. A flat chat capability object would make claims that are true on one route and false on the other.

Absence means "not published", never "supports nothing". What a chat model accepts is covered in Chat, including the parameters that work only on passthrough providers.

Reading it in practice

A client that wants to offer transcription options should ask this route rather than branch on a provider name:

const res = await fetch(`https://api.intfaucet.com/v1/models/${slug}`, {
  headers: { Authorization: `Bearer ${process.env.FAUCET_API_KEY}` },
});
const model = await res.json();

if (model.faucet?.kind === 'transcription') {
  // Only offer formats this gateway will actually bill for.
  showFormats(model.faucet.response_formats);
  // Warn before a parameter silently costs more.
  markSurcharged(model.faucet.surcharged_parameters);
  if (!model.faucet.supports_streaming) disableStreamToggle();
}

A 404 here means the model is not carried, is outside your key's allowlist, or has no usable credential — one answer for all three.