Stanford releases Terminal-Bench-Science; Claude Opus 5 solves only ~30% of tasks
BenBlaiszik · x · 2026-08-28
Stanford researchers released Terminal-Bench-Science v0.1, a benchmark for evaluating AI agents on scientific research workflows across multiple domains. Built by the Terminal-Bench team and global domain experts, the initial version includes 70 tasks. Initial results show that even the top-performing Claude Opus 5 model solves only about 30% of the tasks. The commentary highlights the difficulty of creating robust benchmarks that resist gaming while remaining useful.
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