How Enabling Two Settings Tripled Our Scores on the ARC-AGI-3 Benchmark
OpenAI, Wednesday, July 29th, 2026
OpenAI shows that retaining reasoning and enabling compaction tripled GPT-5.6 performance on ARC-AGI-3.
OpenAI reports that two API settings tripled GPT-5.6 scores on the ARC-AGI-3 benchmark: retaining reasoning across turns and enabling compaction.
The result is notable because it shows how much measured capability depends on harness configuration rather than the model itself.
The post explains what each setting does and why the combination compounds on multi-step reasoning tasks. It also addresses the efficiency implications, since both settings affect token consumption. The finding is directly applicable to developers building agentic workflows on the API.