OpenAI: two Responses API settings tripled its ARC-AGI-3 score
Retained Reasoning and Compaction are live for all Responses API users and cut output tokens 6x — plus the benchmark fine print and a new coding-agent TUI.
Copy markdown
The two switches, in plain terms
Retained Reasoning keeps the model's private chain-of-thought in the conversation instead of throwing it away after each step; Compaction summarizes old context instead of truncating it. Both live on the Responses API — pass the previous response ID and reasoning carries across tool calls and turns.
The jump: 13.3% to 38.3%, at 6x fewer tokens
With both settings on, GPT-5.6 Sol went from 13.3% to 38.3% on the ARC-AGI-3 public set — roughly 3x — while spending 6x fewer output tokens, and edged past Claude Opus 5's official 30.2%. The human baseline sits around 48%.
The catch: not a like-for-like win
ARC Prize says official scores must use a standardized, provider-neutral harness; on a stricter run the same model was reported as low as ~7.8%. These settings favor whoever's API exposes them, so read 38.3% as a best-case tuning number, not a clean head-to-head.
What this means for you
If you run multi-step agents on OpenAI, move off Chat Completions to the Responses API, thread the previous response ID to retain reasoning, and turn on compaction. OpenAI reports the same task landing more accurately and 6x cheaper on output — worth a weekend A/B on your flakiest agent loop.
Elsewhere: one pane for all your coding agents
agent-manager, a new open-source tmux TUI, runs Claude Code, Codex, OpenCode and Grok Build side by side with live status, a group tree, pane preview and CPU/memory gauges. Useful if you're already juggling parallel agent sessions.