Terminal-Bench 4.1 regrades leaderboard: ~1% rewards flip, GLM-5.3 holds
zainhas · x · 2026-11-10
Terminal-Bench 4.1 ships updated verifiers for all tasks, focusing on resisting reward hacking, reducing false negatives, and improving reproducibility. Re-grading the 4.0 leaderboard flipped 1% of rewards, with scores staying within prior confidence intervals.
The regraded chart (max effort) shows Claude Opus 5.5 (claude code) leading at 63% (+0.9%), Sonnet 5.5 +2.1%, GPT-6.1 Sol (codex) +2.7%, GPT-6 Astra +2.1%, Fable 5.1 -0.6%, GLM-5.3 -0.6%, and GPT-5.6 Sol +1.2%. zainhas notes GLM-5.3 "passes the test of time" under the stricter verifiers.
More from coding & agent
- Dev builds a wine-making learning game with Claude Fable 5.2 — 'make a game to learn' — kieranklaassen · 2026-11-10
- Anthropic ships Claude Fable 5.2 with stronger coding, reasoning and verification — lydiahallie · 2026-11-10
- A Perplexity Employee Runs Their Entire Workday on Just 3 AI Teammates — _andychung · 2026-11-10
- After 100B tokens of testing, Berman calls Fable 5.2 the most autonomous model yet — MatthewBerman · 2026-11-10
- First snapshot of official MCP Registry: 57,993 entries, 34.7% with no source — bot2botlabs · 2026-11-10
- Fable 5.2 coordinates 1,000+ subagents for 8 days to build a bootable OS — haider1 · 2026-11-10