Two Responses API settings triple GPT-5.6 Sol's ARC-AGI-3 score
Retained reasoning + compaction lifted Sol from 13.3% to 38.3% with 6x fewer output tokens — past Opus 5, and both settings work in your own agents today.
Copy markdown
The two settings, in plain terms
Retained reasoning keeps the model's private chain-of-thought across tool calls instead of deleting it after every move; compaction summarizes older context instead of truncating it. Both live on OpenAI's Responses API, not the legacy Chat Completions API.
The numbers: 13.3% → 38.3%, at 6x fewer tokens
On the ARC-AGI-3 public set, GPT-5.6 Sol jumped from 13.3% to 38.3% with both settings on — while using roughly 6x fewer output tokens to get there. That edges past Claude Opus 5's 30.2%.
How to switch it on
Move your agent to the Responses API and pass the previous response ID on each call — that alone retains reasoning across tool calls. Then enable compaction to replace rolling truncation. It's configuration, not a new model or fine-tune.
The asterisk: ARC Prize disputes the ranking
ARC Prize argues leaderboards should use a provider-neutral harness; its own run scored Sol at 7.8%. OpenAI counters that a real-world score reflects the model plus its infrastructure, and reportedly ARC tested an older API build missing these features. Treat the '#1' claim with salt — the settings themselves are real and available.
Build this weekend
The takeaway isn't the benchmark — it's that long-horizon agents forget the 'why' behind their own past actions when reasoning gets wiped each turn. Refactor a multi-step agent onto the Responses API with retained reasoning + compaction and watch both accuracy and your token bill move.