The unit of evaluation is now the harness on real work, not the base model alone.
From the Field Notes:
- Introducing FrontierHarness Eval v1.1 (2026-11-19)
- IBM and Kaggle launch Java framework migration leaderboard (2026-10-31)
- SemiAnalysis launches GDPx, a 206-task agentic benchmark across 11 high-value domains; top model scores 12% (2026-11-03)
- Joël Niklaus releases Harness Leaderboard comparing nine coding-agent harnesses on open models (2026-11-07)