OpenAI reports that enabling two API settings tripled GPT-5.6's scores on the ARC-AGI-3 benchmark. The company says the settings work by retaining reasoning and enabling compaction, and that the combination improved both scores and efficiency.
The post highlights the impact of API configuration on benchmark results, showing that how a model is invoked can make a large difference in performance.
This article draws on a single OpenAI post, so the findings are as reported by the company and were not independently verified in the source.