OpenAI demonstrated that enabling retained reasoning and compaction in the Responses API tripled GPT-5.6 Sol's score on ARC-AGI-3 from 13.3% to 38.3% while reducing output tokens by 6x.

Key Details

• GPT-5.6 Sol scored 13.3% with the official generic harness
• Score increased to 38.3% on the public task set when using Responses API harness with retained reasoning and compaction
• Output token usage was reduced by 6x
• GPT-5.5 scored 0.4% on the benchmark

In This Story