invalid_request

The message text is shown in the example below.

When it happens

model is present but not a non-empty string; include_unavailable on GET /v1/models is not true or false; or the model service refused the request as malformed, most often because tool messages do not line up (a tool message whose tool_call_id matches no earlier call).

Example

{
  "error": {
    "type": "invalid_request_error",
    "code": "invalid_request",
    "message": "The request was rejected. Check the message format, tool definitions and parameters.",
    "param": null,
    "request_id": "req_6f1c2a9e0b7d4c3a8e5f1b2d"
  }
}

How to fix it

• Send model as a string from GET /v1/models.
• After a tool call, send exactly one tool message per tool_calls[].id, in the same request that includes the assistant message carrying those calls.
• Do not drop the assistant message that contains tool_calls when you trim history.

Detect it in code

err = resp.json().get("error", {})
if resp.status_code == 400 and err.get("code") == "invalid_request":
    log.warning("%s (%s)", err["message"], err["request_id"])  # fix the request; do not retry unchanged

const { error } = await res.json()
if (res.status === 400 && error?.code === 'invalid_request') {
  console.warn(error.message, error.request_id) // fix the request; do not retry unchanged
}

Related

• All error codes
• Tool results