Research2026-07-29

OpenAI reported that two settings in its Responses API tripled GPT-5.6 Sol's score on ARC-AGI-3, a puzzle-game test of how well AI agents learn on their own. The settings keep the model's private thinking between moves and compress older history rather than deleting it. OpenAI says the result shows benchmark scores partly measure the surrounding test setup, and recommends developers use the same settings it uses in ChatGPT and Codex.

What changed

The benchmark's standard test setup discarded the model's private reasoning after every move and dropped the oldest messages once the conversation grew too long.

What it unlocks

Running long, multi-step agent tasks where the model keeps its own earlier thinking and a compressed history instead of starting over each turn.

  • 13.3% to 38.3% on the ARC-AGI-3 public set
  • 6x fewer output tokens
  • human testers averaged an estimated 48%
  • 175,000 token context limit

What you need to act on it

  • use of OpenAI's Responses API rather than the older Chat Completions API
  • passing the previous response ID to retain reasoning
  • compaction enabled

Send this to someone who needs it

Shares the story and its sources. Nothing about you.

What does this mean for your job?

This is the story as everyone gets it. Once a week we send you the version written for your role — what changed, why it matters for the work you actually do, and one thing to try. Free while we tune it.