Research

Publications

Studies, not blog posts: a question, a method, the numbers, and the sources — published in full. The living tables live on Rankings; these are the dated snapshots with the reasoning attached.

2026-11-15

Which model is best at what

Publication 001 — the per-task leaders across twelve categories: writing, reasoning, agentic coding, code review, research accuracy, everyday chat, image generation, long context, overall intelligence, value for money, and routing itself.

Opus 5 · 96.0% SWE-bench GPQA boards disagree Decision Index inside Model rankings

Next in the program: routing accuracy under real chat workloads, and the decision-model follow-up once independent re-runs of the Decision Index land.