--- format: "aidr-story-markdown/v1" id: "e2c32579ca3943549b495ba7cfccbe66e39dcfee2221b05ff75b5fe93ac63f42" canonical_url: "https://aidr.today/e2c32579?lang=en" title: "Claude Opus 5 Solves Only 30% of New Scientific AI Benchmark" lang: "en" requested_lang: "en" available_langs: ["en","vi"] translation_fallback: null fallback_fields: [] published_at: "2026-08-27T20:55:33.000Z" category: "Research" topics: ["anthropic","claude","benchmark","agent"] source_urls: ["https://huggingnews.com/ai/claude-opus-5-solves-only-30percent-of-new-scientific-ai-benchmark-d3363d27","https://x.com/modal/status/2093046013640450225","https://x.com/madiator/status/2093054302634062216"] summary: "A Stanford-led community effort released Terminal-Bench-Science v0.1 to evaluate AI agents on research workflows across scientific domains. The initial benchmark contains 70 tasks, and Claude Opus 5 solved approximately 30% of them. The team behind the original Terminal-Bench collaborated with scientific experts from research institutions worldwide and received sponsorship from Bespoke Labs. The initiative aims to develop reliable AI partners that allow scientists to prioritize human-centric work." --- # Claude Opus 5 Solves Only 30% of New Scientific AI Benchmark > [Open the canonical story]() **Published:** 2026-08-27T20:55:33.000Z **Category:** Research **Topics:** anthropic, claude, benchmark, agent ## Summary A Stanford\-led community effort released Terminal\-Bench\-Science v0\.1 to evaluate AI agents on research workflows across scientific domains\. The initial benchmark contains 70 tasks, and Claude Opus 5 solved approximately 30% of them\. The team behind the original Terminal\-Bench collaborated with scientific experts from research institutions worldwide and received sponsorship from Bespoke Labs\. The initiative aims to develop reliable AI partners that allow scientists to prioritize human\-centric work\. ## Sources - [Story source]() - [Story source]() - [Supporting source]()