Stanford-Backed Team Launches Terminal-Bench-Science, an AI-for-Science Agent Benchmark
ArtificialAnlys · x · 2026-09-25
The Terminal-Bench team and Stanford researchers launched Terminal-Bench-Science, extending agent benchmarking to real scientific research workflows across five domains with programmatic verification in containerized environments. Unlike static science benchmarks, it ships regular frontier-calibrated releases and forms a contribution loop with labs; frontier models were benchmarked on it within a week, appearing in Claude Fable 5.1 and GPT-6 Astra release notes. Artificial Analysis runs all models with a unified mini-swe-agent harness; contributions for 0.2 are open until October 5.
Related event: Stanford Releases Terminal-Bench-Science, Best Models Solve Only 30%(2 posts)→
More from Models
- Commit leak reveals unreleased Gemini 4 variants: Argon, Hydrogen and Helix (unconfirmed) — BlackHC · 2026-11-11
- Rewiring just 89 of 442,368 neurons in Qwen3-8B matches full SFT for a persona — maxsloef · 2026-11-11
- Unsloth ships Qwen3.8-27B Dynamic V3.1 GGUFs with 5-10% accuracy gains — danielhanchen · 2026-11-11
- Zvi: being forced back to an older model is the best reminder of AI progress — TheZvi · 2026-11-11
- Screenshots Made a 4B Browser Agent Worse: A Postmortem With Real Numbers — Stunning-Sherbet1853 · 2026-11-11
- Codex safety filter flags developer's hobby git forge app as cybersecurity risk — brianmichel · 2026-11-11