OpenAI released GPT-6.1 Terra in the API on 26 October 2026. If you move a document workflow to the new model, check the response status before treating its text as a finished file. A model change does not remove your application’s output cap.
A polished final sentence is not a reliable completion signal. In the OpenAI Responses API, inspect status and, when present, incomplete_details before accepting a generated document.
Separate the available space
An output allowance can include reasoning tokens as well as visible text. A response may therefore reach max_output_tokens before producing the expected answer. The context window is a separate constraint on what can fit in the request and response together.
Suppose you request a twelve-section guide and receive seven sections. Record which sections exist, then request only the missing material with enough source context. “Continue” alone can repeat paragraphs or change terminology.
Accept only a complete artifact
Check the promised section list, closing delimiters and actual response status. Parse structured output before saving it. Increasing a cap may increase cost and does not repair unsupported facts. Keep partial output separate from the version that readers or downstream tools receive.
If status is incomplete and incomplete_details.reason is max_output_tokens, preserve the partial result for inspection. Do not mark it successful merely because the HTTP request succeeded. Increase the cap only within the selected model’s supported limits and your budget, or divide the task into smaller sections; a fresh attempt can incur additional usage.