FULL STORY
Stanford's Science Agent Benchmark Stumps Top Models
Stanford released Terminal-Bench-Science 0.1, extending agentic benchmarks to scientific tasks. GPT-6 Astra and Opus 5.5 top the new leaderboard, yet even the best solve only about a third of tasks.
2026-09-23 ~ 2026-09-25 · 2 episodes · 6 posts
Episode 1 · Stanford Releases Terminal-Bench-Science, Best Models Solve Only 30% (2026-09-23, 2 posts)
Stanford and global research institutions released Terminal-Bench-Science 0.1, extending Terminal-Bench's evaluation paradigm to scientific research with 70 tasks. The best model solves only about 30%, highlighting the gap in scientific agent capabilities.
- Stanford-led Terminal-Bench-Science debuts: top model solves just 30% of 70 expert tasks — geoffwolfe · 2026-09-23
- Stanford-Backed Team Launches Terminal-Bench-Science, an AI-for-Science Agent Benchmark — ArtificialAnlys · 2026-09-25
Episode 2 · Terminal-Bench-Science Leaderboard Debuts with GPT-6 Astra and Claude Opus 5.5 on Top (2026-09-25, 4 posts)
Artificial Analysis launched Terminal-Bench-Science 0.1, a 70-task agentic science benchmark where GPT-6 Astra and Claude Opus 5.5 lead by 20 points and are the only models scoring above 50%.
- Claude Opus 5.5 Gains 38 Points on Terminal-Bench-Science at 5x the Cost — ArtificialAnlys · 2026-09-25
- GPT-6 Astra and Opus 5.5 Lead Terminal-Bench-Science by ~20 Points — ArtificialAnlys · 2026-09-25
- Terminal-Bench-Science leaderboard launches with GPT-6 Astra at 63% — scaling01 · 2026-09-25
- Terminal-Bench-Science leaderboard launches: GPT-6 Astra 63%, Opus 5.5 62%, open weights ~10% — CurieuxExplorer · 2026-09-25