Supported parameters

The wire is iso-OpenAI and the parameter surface is a documented subset. What is honored, what is accepted and inert, and what is a 400 naming itself.

chat.completions.create takes OpenAI's own request type, so TypeScript will let you set any field OpenAI declares. What the Fleeex proxy accepts inside that request is narrower: a large subset, but a subset. This page is the list.

It fails loudly rather than quietly. One rule decides which way a parameter goes:

A parameter that would change the answer, the billing, or a guarantee Fleeex would be implying is a 400 naming that parameter, because silently dropping it would hand you a wrong answer you could not detect. A parameter that changes nothing observable is accepted and ignored, because a 400 there would break an integration for no benefit.

Honored

Forwarded to the model, and billed accordingly.

Parameter	Notes
`model`	A catalog id or an alias. See Choosing a model.
`messages`	Roles `system`, `user`, `assistant`, `tool`. `content` is a string or an array of `text` / `image_url` parts. Includes `tool_call_id` on a `tool` message and `tool_calls` replayed on an assistant message.
`tools`, `tool_choice`	`tool_choice` is `'auto'`, `'required'`, or `{ type: 'function', function: { name } }`. See Function calling.
`response_format`	The `json_schema` form. See Structured output.
`max_completion_tokens` / `max_tokens`	Integer 1–32,768. Defaults to `1024`. See below.
`temperature`	0–2.
`top_p`	0–1.
`stop`	A string, or up to 4 strings.
`stream`	`true` returns an SSE stream of iso-OpenAI chunks. See Streaming.
`stream_options.include_usage`	Adds the trailing usage chunk. Requires `stream: true`. See Streaming usage.

Accepted and inert

Validated, then ignored, because they change nothing observable. The OpenAI SDK, LangChain, LiteLLM and the Vercel AI SDK all send several of these by default, which is why rejecting them would cost an integration and buy nothing.

Parameter	Accepted at	Why it changes nothing
`user`	any string	The billed pair is the verified API key plus `x-fleeex-user`. A body field must never designate a user.
`metadata`	any bounded record	Not forwarded, not stored.
`store`	`true` or `false`	Fleeex retains no prompt or completion content and exposes no endpoint that could return one, so there is nothing for `true` to change.
`n`	`1` only	One completion per request.
`logit_bias`	`{}` only
`frequency_penalty`	`0` only	OpenAI's default, and the provider's behavior.
`presence_penalty`	`0` only	Same.
`logprobs`	`false` only
`response_format`	`{ type: 'text' }`	OpenAI's default asks for nothing, so nothing is forwarded.
`image_url.detail`	`'auto'` only	OpenAI's default, and decorative.
`json_schema.strict`	`true` or `false`	The provider constrains the output unconditionally, since it has no non-strict mode, so `strict: false` gets a stronger guarantee than OpenAI promises, never a weaker one.

A field that is decorative only at one value pins that value rather than being waved through: n: 1 is decorative, n: 2 asks for two completions that are not produced, and returning one would be both a wrong answer and a billing question.

An explicit null reads as "unset" on every optional field. That is how several OpenAI clients serialize a field they did not set, so it is normalized away rather than forwarded.

Rejected: a `400` naming the parameter

Parameter	Why
an unsupported or disabled `model`	Not in the catalog. The message echoes the value you sent, not what an alias resolved to.
`seed`	The provider's inference config has no seed and no portable equivalent. Accepting it silently would advertise a reproducibility Fleeex cannot deliver.
`n` greater than 1	Exactly one completion is produced per request.
a non-empty `logit_bias`	Would change the output.
a non-zero `frequency_penalty` / `presence_penalty`	Same.
`logprobs: true`	Not produced.
`tool_choice: 'none'`	The provider has no way to expose the tools while forbidding their use. Dropping the tools would change both the answer and the input tokens billed; ignoring the field would let the model call a tool you forbade. Send the request without `tools` instead.
`tool_choice` without `tools`	Nothing to choose from.
`function.strict`, and any undeclared sub-field of a tool	Not implemented; see `json_schema.strict` above, which is a different field.
`response_format: { type: 'json_object' }`	There is no schema-less JSON mode to map onto, and every approximation is undetectably dishonest. Send a `json_schema`, as the message says.
an `image_url.url` that is not a `data:` URI	Fleeex would have to fetch it server-side on the billing path. Send `data:image/png;base64,…`. See Vision.
an `image_url.detail` other than `'auto'`	`low` and `high` select a different image tokenization: they would change the answer and the tokens billed.
`stream_options` without `stream: true`	The parameter only exists in the streaming wire.
anything undeclared	Including an unknown key nested inside `stream_options`, a tool, or `json_schema`.

The message names every rejected parameter and lists the accepted ones, read off the request schema itself, so none of this has to be discovered by trial and error:

400 response

{
  "error": {
    "code": "BAD_REQUEST",
    "message": "Unsupported parameter: 'seed'. Accepted: model, messages, max_completion_tokens, …",
    "correlationId": "…"
  }
}

`max_completion_tokens` sizes the reservation

max_completion_tokens is OpenAI's current name and max_tokens the deprecated one. Both are accepted, and the new name wins when you send both. Neither is only a cap on the model: this number also sizes the wallet reservation Fleeex places before calling the provider, so a needlessly large value can turn into a 402 on a wallet that would have covered the real answer. It defaults to 1024.

Bounds

Size limits, so no single field can be arbitrarily large. All of them are a 400.

	Bound
Request body	1 MB (a larger body is a `413`), nesting at most 32 levels deep
`messages`	1–200 entries
`content`	256,000 characters per message, or per text part; at most 20 parts
Images	at most 8 per request, each decoding to at most 512 KB. See Vision
`tools`	at most 128; `name` at most 64 characters on `[A-Za-z0-9_-]`; `description` at most 4,096; `parameters` a JSON object, at most 16,384 serialized characters and 10 levels deep
`tool_calls`	at most 32 per message; each `arguments` at most 32,768 characters; `tool_call_id` at most 128
`json_schema`	`name` at most 64 characters on `[A-Za-z0-9_-]`; `schema` a JSON object, at most 16,384 serialized characters and 10 levels deep; `description` at most 4,096

The always-current version

This page tracks the route's own OpenAPI description, POST /v1/chat/completions, which is generated from the request schema and is therefore authoritative. If the two ever disagree, the route wins.