Vibitely Daily

Sunday, 22 November 2026, story 1 of 7

OpenAI's GPT-6.2 API can drop reasoning when history is edited

The Responses API also drops temperature and top_p, and OpenAI's new Astra and Sol prompt templates target terse answers and needless questions.

Illustration: A spool of thread with a few strands snipped mid-length, the cut ends lying loose beside a needle on a plain surface.

What happened

OpenAI says that, starting with GPT-6.2, the Responses API discards earlier reasoning content if developers edit conversation history, switch to certain other models, or change the end user or organization. The company says this can hurt response quality and raise token use as the model repeats work. Temperature and top_p are no longer supported. OpenAI recommends medium reasoning effort for new integrations, and setups on none or minimal should move to low.

OpenAI also published copy-ready system prompts for GPT-6.2 Astra and Sol, which launched November 17. They address shorter default answers, unnecessary questions and coding habits such as writing new tests, The Decoder reports.

What it means

Any agent that trims, summarizes or rewrites its history to manage context may now erase the model's reasoning each time it does so. The same goes for code that swaps models mid-session or changes the end user or organization between calls. The source does not say how large the quality loss or token increase is, so teams will have to measure it on their own workloads.

Cost is part of the picture. Astra's input price rises from $8 to $16 per million tokens beyond 272,000 tokens, so repeated reasoning in long sessions could push bills up faster. Code that sets temperature or top_p needs a check, though the source does not say how the API responds to those parameters. The prompt templates are OpenAI's own fixes for behavior it observed, not independently tested.

What to do: Audit agent code that edits history, swaps models or changes users mid-session, and test removing temperature and top_p before moving to GPT-6.2.

Source

Filed under Models. Published .

"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.