← Daily

Field note · No. 034 · November 19, 2026

GPT-6.2 drops reasoning when your agent edits its own history

GPT-6.2 Astra and Sol change how the Responses API treats reasoning items and the top-level instructions field. Harnesses that compact history, switch models mid-session or keep their system prompt in instructions will lose reasoning context and cache hits without a single error.

What changed

OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol to the API on November 17, 2026. The changelog positions Astra for demanding reasoning and coding tasks and Sol for everyday coding and agent workflows. Both support the Responses API and Chat Completions, but tool calling requires the Responses API.

The model choice is not the important part. The same changelog entry says GPT-6.2 changes how the Responses API preserves reasoning and how it handles the top-level instructions field. The migration guide spells out three changes:

  • Reasoning items are dropped after mid-conversation edits to the history (prefix edits), after certain model changes including downgrades, and after changes to the end user or organization.
  • Top-level instructions are now appended after the input. Reusable instructions only stay eligible for prompt caching if they sit in a developer message as the first item in input.
  • Requests that set reasoning.effort or reasoning_effort to none are rejected. The guide recommends starting at medium.

Separately, OpenAI announced background mode targets of 2h or 8h for the first response, set with completion_window, at up to 40% less per token than Standard. These windows are in public beta.

Why it matters for builders

Most agent harnesses edit their own prefix as a matter of routine. Compaction summarizes old turns, trimming drops stale tool output, and retries rewrite the last step. Under GPT-6.2, each of these discards the reasoning the model had built up. Routers that downgrade to a cheaper model mid-session, and multi-tenant systems that change the end user on a shared conversation, hit the same rule.

None of this raises an error. The agent keeps running, just with less context than before and with earlier input no longer eligible for caching. You notice it as worse multi-step behaviour and a higher bill, which is the hardest kind of regression to trace back to a model upgrade.

The instructions change is quieter still. A stable system prompt kept in the top-level field now lands after the input, so it stops being a cacheable prefix. Nothing breaks, the cache just stops working.

The effort: none rejection is the only loud change, and it is the easiest to fix.

What I would do with it

I would treat the conversation prefix as append-only by default and make every edit to it a deliberate, logged decision. Concretely:

  • Move the reusable system prompt into a leading developer message in input. Keep top-level instructions for guidance that applies only to the current request, such as a timestamp or a per-turn hint.
  • Compact at explicit boundaries, like the end of a plan stage, rather than on every turn, and accept that reasoning resets there.
  • Pick the model at session start and pin it. If a router must downgrade mid-session, record that as a reasoning reset rather than a free cost saving.
  • Search the codebase for none effort settings and replace them with medium, then tune per task from measurements.
  • Track cache hits per turn before and after the migration, and fail the rollout if they drop.

For evals, overnight batch work and anything that does not need an answer now, I would trial the background completion windows, keeping in mind they are still in beta.

When a model ties reasoning and caching to an unchanged prefix, every history rewrite in your harness becomes a cost and quality decision, so make those rewrites rare, explicit and measured.

Sources