GPT-5.6 Sol dethrones Opus 5 on ARC-AGI-3

openai claims its gpt-5.6 sol model reached 38.3% accuracy on the arc-agi-3 benchmark, overtaking anthropic's opus 5 which recently set a record at 30.2%. however, this high score depends on using openai's responses api with 'retained reasoning' and 'compaction' settings, whereas the standard official harness restricted the model to 7.8% because it discarded reasoning between steps. arc prize co-founder françois chollet acknowledged that general-purpose api settings are fair game, admitting the previous official harness may have disadvantaged openai. the dispute underscores that ai benchmarks increasingly measure the technical setup surrounding the model, not just the base intelligence.