Skip to main content
2026/9/18 14:41 (UTC+8) · Docs Update · OpenAI 📚 OpenAI models refuse with a 200, no error and no category, which shows up as a JSON parsing failure in translation tasks When a GPT model receives a request that violates the provider’s usage policy, it returns HTTP 200, finish_reason: "stop" and a short refusal. There is no error code and no refusal category, and the call is billed as normal output. A task that expects JSON gets a plain sentence instead, so the client reports a parsing error such as Unrecognized token. APIYI has enabled automatic content safety failover for non-streaming requests: when one official route triggers a content filter, the request is retried on another official route. Streaming requests cannot be switched once output has started. The new FAQ shows the refusal response body and how to detect it, and recommends archiving failed cases internally for analysis before retrying that content with another model. See What Does an OpenAI Model Refusal Look Like?.
← Back to Live · 📚 Monthly archive