I’m getting HTTP 422 “Unprocessable Entity.” What does it mean?

Last updated: August 6, 2026

422 means the request reached us but something in it is invalid. Unlike some providers, we validate your request — including tool/function JSON schemas — rather than silently ignoring unsupported fields. The message in the response body tells you the cause; here are the common ones:

A. Tool schema issues

Message	Cause	Fix
`'oneOf' with multiple schemas is not supported`	A tool/function schema uses `oneOf`, often injected by an MCP server or agent framework.	Replace `oneOf` with `anyOf` in the tool schema.
`Reference not found: #/…`	A tool schema contains a JSON `$ref` whose definition wasn’t included — common when auto-generated from an OpenAPI spec.	Inline the referenced definitions so the schema is self-contained.
`tool calling is not supported`	The model doesn’t have tool calling enabled on that endpoint. LangChain’s `bind_tools` sends the same field and fails the same way.	Use a tool-capable model — check the “Supported Functionality” panel via `GET /v1/models`.
`Function names should not be blank.`	A tool object is missing its `function` sub-field, or the `function` object has no `name`.	Include at least a `name` field in the `function` object.

B. Field combination conflicts

Message	Cause	Fix
`The 'response_format' field cannot be set when tools are specified`	`response_format` and `tools` were sent together.	Remove `response_format` when using tools, or use them in separate requests.
`The field 'min_tokens' is unsupported when tools are specified`	`min_tokens` and `tools` were sent together.	Remove `min_tokens` when using tools.

C. Unknown fields — we reject unknown fields rather than silently ignoring them:

{"message":"invalid JSON payload: no such field: '<field_name>'"}

Common fields that trigger this: thinking, advanced, enable_thinking, extended_thinking. If you need to toggle reasoning on a model that supports it, use chat_template_kwargs: {"enable_thinking": true} instead.

reasoning_effort sent as a boolean (e.g., True) returns a different message: "expected number or string at reasoning_effort". Use a string value ("high", "medium", "low") or a number instead.

D. Malformed messages — sending messages as a plain string instead of an array of objects:

{"message":"invalid JSON payload: unexpected character: '\"'; expected '{' at messages"}

Sending the messages field to the /v1/completions endpoint (which expects prompt) returns "no such field: 'messages'". Chat requests should go to /v1/chat/completions.

E. Parameter range validation

Message	Cause	Fix
`top_p should be in range between 0 (exclusive) and 1.0.`	`top_p` set > 1.0	Set to a value in (0, 1.0]
`max_tokens/max_completion_tokens must be greater than 0`	`max_tokens` ≤ 0	Set to a positive integer
`n should be <= 32.`	`n` exceeds 32	Reduce to 32 or fewer
`no valid input; 'messages' parameter is required.`	`messages` is empty or missing	Provide at least one message object

F. top_logprobs — top_logprobs is not supported on any current serverless model because all models have reasoning content parsing enabled:

{"message":"top_logprobs is not supported with reasoning content parsing."}

logprobs: true alone (without top_logprobs) is accepted.

G. Context length and KV-cache

Message	Cause	Fix
`input is too long (received … tokens; max …)`	Input exceeds the model’s context window. Context length = input + output combined.	Shorten the prompt or switch to a model with a larger context window.
`not enough kv_cache_size to process this request; try with smaller max_tokens/n option or shorter prompt`	Prompt + max_tokens × n exceeds available KV-cache capacity at request time, often under load.	Reduce `max_tokens`/`n`, or shorten the prompt, and retry.