The model at the top of a composite index need not lead every category. Epoch AI's September breakdown puts OpenAI's GPT-6 Astra ahead overall and in mathematics, while Anthropic's Claude Fable 5.1 has a higher software-engineering point estimate. Source.
The category estimates rely on relatively few benchmarks and have overlapping uncertainty intervals. Astra's overall score also changed as more results arrived. The breakdown is useful for choosing what to investigate for a workload, rather than treating a small leaderboard gap as a settled universal ordering.
