Service update · OpenAI
GPT-6.2 reasoning in the Responses API is now tied to earlier context, and editing it drops later reasoning
- Announced
Starting with GPT-6.2, each reasoning item is tied to the conversation history before it, so editing earlier messages makes the API drop the reasoning that follows the change; the model may then redo reasoning, adding tokens, cost and latency. OpenAI recommends keeping history unchanged and appending new turns. Changing reasoning effort or adding tools can also edit the prefix by accident; the Responses API's new input_transformations object and the Usage dashboard's new 'Reasoning removals' section help detect it.