Chat Completions

POST https://api.airouter.io/v1/chat/completions is a drop-in replacement for the OpenAI chat completions endpoint (/chat/completions without the /v1 prefix works as well). Send your AI Router API key as a bearer token:

Everything the OpenAI API supports (streaming, tool calling, response formats, temperature, max_tokens, ...) is passed through to the selected model. On top of that, AI Router understands the following parameters. With the OpenAI SDKs, put them into extra_body; with LangChain, into model_kwargs.extra_body.

Request parameters

Parameter	Type	Default	Description
`model`	string	team default	A concrete model name pins the request to that model. `auto` (or omitting it while the team default is `auto`) lets AI Router decide. See Default Model.
`models`	array of strings	team's enabled models	Candidate allowlist for this request. A single model turns the request into a plain gateway call. See Choosing Models.
`weighting`	object	team preferences	`quality`, `costs` and `latency` boost factors, each a non-negative number. See Weighting.
`model_routing`	boolean	`true`	`false` returns the selected model name instead of executing the request (model selection mode, see below).
`embedding`	array of numbers	-	Full privacy mode: the embedding of the prompt instead of `messages`. Implies `model_routing: false`.
`embedding_type`	string	`paraphrase-multilingual-mpnet-base-v2`	Embedding model of `embedding`: `text-embedding-3-small` or the FastEmbed default.
`max_tokens`	integer	-	Also used as the expected completion length for the cost and latency prediction.
`tools`	array	-	Standard OpenAI tool definitions. When present, only models with tool support are considered.
`reasoning_effort`	string	-	Passed through and clamped to the tiers the selected model supports.
`expose_predicted_savings`	boolean	`false`	Model selection mode only: append the predicted cost and latency savings to the response.
`expose_decision_id`	boolean	`false`	Model selection mode only: append the routing decision id for usage reporting.

Per-request parameters override the team configuration for that request; API key overrides override the team configuration for every request of that key.

Model routing (default)

The response is a standard chat completion object. model in the response names the model that produced the answer. When the selected model fails, AI Router automatically falls back to the next best candidates of the same decision; with connected provider keys the fallback chain stays on your own keys.

Model selection (`model_routing: false`)

With model_routing: false AI Router returns the selected model name as the message content instead of calling the model:

The content of choices[0].message.content is comma-separated:

Options	Content
none	`<selected_model>`
`expose_decision_id`	`<selected_model>,<decision_id>`
`expose_predicted_savings`	`<selected_model>,<cost_savings>,<latency_savings>`
both	`<selected_model>,<cost_savings>,<latency_savings>,<decision_id>`

The savings are fractions relative to AI Router's reference model (0.41 means 41% lower predicted cost). This is the legacy selection path used by the Python and TypeScript SDKs; new integrations should prefer the Decision API, which returns a structured, ranked response.

Errors

Status	Meaning
`400`	Invalid parameters, for example unknown `weighting` keys.
`401`	Missing or invalid API key.
`403`	The feature is not included in your plan (model routing, full privacy).
`503`	No usable provider key is available for the selected models. Connect a provider key.

The OpenAI-compatible endpoint and the AI Router specific request parameters

Request parameters

Model routing (default)

Model selection (model_routing: false)

Errors

Model selection (`model_routing: false`)