Stanford's Terminal-Bench-Science: Claude Opus 5 solves only ~30% of research tasks
realsohamparekh · x · 2026-08-28
A Stanford-led community effort released Terminal-Bench-Science (v0.1, 70 tasks), a benchmark evaluating AI agents on research workflows across scientific domains, built by the Terminal-Bench team with domain experts at research institutions worldwide. Claude Opus 5 solves only 30%.
One user broke down its task design philosophy: it explicitly rejects textbook differential equations, generic CSV cleaning, standard undergrad ML exercises, literature summaries, and tasks solvable by looking up one obscure fact (an interesting exclusion). As a 0.1.0 release, its future looks promising.
Related event: Stanford-led Terminal-Bench-Science Debuts; Claude Opus 5 Passes Just 30%(19 posts)→
More from Research
- Robotics papers on arXiv have grown ~6x in five years, peaking at 676 in a single week — Rewkang · 2026-10-15
- MicroSDF: microfacet-driven neural SDFs for mixed-reflectance 3D reconstruction — ssh4net · 2026-10-15
- 'Thinking leakage': no-think RL gains in hybrid models just reuse think-mode skills — kaust-generative-ai · 2026-10-15
- COMPASS predicts model-pair compatibility before alignment or merging — epfl-ch · 2026-10-15
- BERT tokenizer gets 50x speedup by skipping text normalization, not optimizing it — LysandreJik · 2026-10-15
- LLM agents autonomously design a SAR ADC on SKY130 from schematic to signoff, paper finds — IgorCarron · 2026-10-15