AI Aulis

AI NEWS IN ENGLISH · 20 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

41 stories · next issue tomorrow morning · How the paper is made →

← Issue 20 November 2026Paper

OpenAI: Responses API drops GPT-6.2 reasoning after an edited message

AI-generated contentWritten by AI

x.com, developers.openai.com · 4 sources · Sources published: 20 November 2026 · Some source publication dates are unknown

OpenAI advises developers using GPT-6.2 Astra and Sol to keep conversation history unchanged and append new turns. Preserved reasoning gives better results with fewer tokens, the company said.

The OpenAI Developers account announced a change on X on 19 November that affects the company’s newest frontier models, starting with GPT-6.2. Each reasoning item is now tied to the conversation history before it. If a developer edits an earlier message, the Responses API drops the reasoning that follows the edited point in the history.

A reasoning model spends tokens on hidden reasoning before it answers. The Responses API can store this reasoning and bring it into the context of the next turn, so the model continues from its earlier work instead of redoing it. OpenAI’s guide says the reasoning text is still not shown to the developer.

In its GPT-6.2 guide, OpenAI lists three situations in which reasoning is lost: the conversation history is edited mid-conversation; the model is switched in certain ways, for example to a lower-tier model; or the end user or organisation changes. Answer quality can then decline, the company said, and the model may use more tokens repeating earlier work.

The change matters most for the harness, the program that drives an agent, gives the model its tools and runs its steps. If the harness alters earlier developer messages or tool calls, for example, the reasoning is lost. OpenAI advises appending new turns to the end of the history.

The same rule applies to the cache, which reuses the start of an input if it exactly matches input processed earlier. GPT-6.2 appends the instructions in the top-level instructions field after the input. OpenAI therefore recommends putting repeated instructions in a developer message as the first item in the input so that the cache can pick them up. Developers migrating from GPT-5.5 or earlier models must also replace the prompt_cache_retention setting with prompt_cache_options.ttl, set to "30m".

OpenAI says preserved reasoning brings better results, fewer tokens and more cache hits. The company gives no figures to support the claim.

Reasoning is preserved when a developer references the previous response with the previous_response_id identifier, or sends the full history in its original order and appends new items. If responses are not stored with OpenAI, the encrypted reasoning items must be included. Reasoning can be reused only within the same model family.

Report an error in this story