Models and offers
Find a callable model ID and understand why offers differ.
Find the model ID
Open Marketplace, choose a model, and copy its ID. Use the ID in your request's model field; the display name is for reading.
Models use canonical, vendor-neutral IDs such as claude-sonnet-5-5, rather than anthropic/claude-sonnet-5-5. Previously registered qualified IDs remain aliases. Copy the published ID instead of constructing a provider prefix.
List models available to your key
Use the environment variables from Quickstart.
curl --fail-with-body "$UNUSED_BASE_URL/models" \
-H "Authorization: Bearer $UNUSED_API_KEY"The default response has object: "list" and a data array with model id values. Sending an anthropic-version header selects the Anthropic-shaped list, with data, has_more, first_id, and last_id.
This is a filtered availability list. Key restrictions, routing model lists, the price cap, offer health, and usable inventory can remove a model. Your own offers are excluded from your buyer routes.
Compare offers for the same model
An offer keeps its own ID, input/output rates, cache policy, available inventory, and status. A model's displayed best price is not a promise that every request will use that offer. Your settings, requested context/output size, and recent failures affect eligibility.
TTFT means time to first generated token, including reasoning and tool calls. Marketplace's model row uses the TTFT of the offer supplying that row's displayed best price. Open the model's offers to compare them; a missing TTFT means no current measurement, not zero latency.
Liquidity (USD) describes reported or estimated capacity value. It is separate from your credit balance. Subscription liquidity is an estimated API-equivalent value, not cash or a guaranteed quota; check the Estimated API value, Partial data, and Last reported labels.
Check capabilities before choosing an agent
The gateway accepts OpenAI Chat Completions and Anthropic Messages. Supported conversions include system instructions, images, tool calls/results, reasoning, structured JSON, stop reasons, and usage. This does not add capabilities a model or adapter lacks.
| Offer type or request | Limitation to check |
|---|---|
| Context and output | The estimated input plus requested output must fit an eligible offer's context window; output must fit its output ceiling. |
| OpenAI request options | n must be 1; logprobs, top_logprobs, logit_bias, and audio are rejected. modalities supports text only. |
| Claude web session | Text only, with estimated usage; tools, images, thinking/cache controls, sampling controls, stop sequences, and structured output are unsupported. |
| OpenAI subscription adapter | Stop sequences and top_k are unsupported. Client requests still use Chat Completions or Messages. |
/v1/responses is not a public endpoint. Choose a client that can use one of the two supported inference formats; see coding agents.