News
November 22, 2026 · OpenAI

OpenAI updates its prompting advice for GPT-6.2 Astra and Sol developers

OpenAI's GPT-6.2 docs now include prompt templates for terser default answers, fewer confirmation pauses and leaner code, plus API changes that can discard earlier reasoning.

OpenAI has published new prompting advice for developers moving to GPT-6.2 Astra and GPT-6.2 Sol. The guidance sits on the Using GPT-6.2 page of its API documentation, and The Decoder reported on it on November 21. OpenAI says the examples target behaviours it has seen in both models, and it offers ready-made system prompt snippets for each one.

What's new

  • Shorter answers by default: according to the docs, GPT-6.2 Astra writes more concisely and directly than earlier models, so OpenAI suggests revisiting older instructions that tell the model to keep things short.
  • Fewer check-ins: OpenAI says that if a request is ambiguous, the model can stop to suggest what it would do next and then wait for a reply. Its sample prompt tells the model to keep going on authorised work, state its assumptions, and stop only for a change in scope, a real blocker, a user decision or a destructive action.
  • Coding conventions: the coding template asks the model to reuse existing logic, avoid needless helper layers and add tests only when asked, when the repo requires them, or to catch a specific regression.
  • Frontend work: OpenAI recommends naming the imagery and typography you want and cutting decorative filler from interfaces.
  • API changes: per the docs, the Responses API now drops earlier reasoning items after edits to the conversation history, some model switches, or a change of end user or organisation, which OpenAI says can lower answer quality and raise token use.
  • Parameters: migrating requests should omit temperature and top_p, and OpenAI suggests starting new integrations at medium reasoning effort.

Why it matters

Analysis: prompts written for older models can now work against you. An instruction to be brief may make already terse answers too thin, and a cautious prompt may cause more pauses for confirmation. Teams building on the API, or using ChatGPT and Codex in agent-style workflows, should retest their system prompts after switching. The change to reasoning items matters for cost: apps that rewrite earlier messages in a conversation may pay for repeated work without noticing.

Source: OpenAI