I'm seeing stray tokens (e.g. `<turn|>`) or malformed JSON in the output.

This is almost always a chat-template or model-specific formatting issue rather than an outage. First confirm you're calling the correct model ID, and for structured output, that your request uses the documented JSON/schema options. If a stray token persists for one specific model, send our team a sample request/response payload and let us know whether a custom chat template is configured — we'll investigate.

For structured output, include an example of the desired output in your prompt rather than relying on the response_format schema alone. A good prompt matters as much as the schema; both contribute to reliable structured output.

See our blog for more information.