Leaked internal benchmarks from OpenAI suggest GPT-6 achieves near-human performance on a battery of reasoning tasks that its predecessor struggled with. The numbers, if accurate, represent a step-change rather than an incremental improvement.
What the Benchmarks Show
According to documents circulating in AI research circles, GPT-6 scores 94% on the ARC-AGI benchmark — a test specifically designed to resist pattern-matching by requiring genuine reasoning. GPT-4 scored 33%. GPT-4o scored 53%.
