Summary & Key Takeaways ​

• OpenAI achieved a threefold increase in GPT-5.6's ARC-AGI-3 benchmark scores.
• This improvement was due to enabling two specific API settings.
• The settings enhanced performance by retaining reasoning capabilities.
• They also improved efficiency through a compaction mechanism.

Our Commentary ​

This is a classic OpenAI move, isn't it? A quiet blog post about a massive performance jump. "Tripled scores" sounds huge, but it's always about the specific benchmark. I wonder what these "two settings" actually are. It feels like a peek behind the curtain, but only just.