Stanford-led Terminal-Bench-Science Debuts; Claude Opus 5 Passes Just 30%

A Stanford-led community effort released Terminal-Bench-Science (TBS) v0.1, a benchmark for evaluating AI agents on interdisciplinary scientific workflows. The strongest model, Claude Opus 5, passes only 30.0% of tasks, showing a large gap between frontier models and real research automation and providing a new yardstick for tracking progress.

Confirmed

• v0.1 contains 70 tasks spanning life sciences, physics, math, engineering and earth science; each task was derived from a researcher's own computational workflow and passed multi-stage expert review
• 70 tasks were selected from 920 proposals via three review layers (domain reviewer, technical reviewer, Bar Raiser); authors estimate over 80 hours of work per task
• Claude Opus 5 (Claude Code) leads overall at 30.0%, ahead of GPT-5.6 Sol, which wins in math at 31.4%
• Compared with Terminal-Bench 3.0, TBS offers comparable discrimination among frontier models while lowering all pass rates by over 10 points (Opus 5: 43% → 30%; Fable 5 scores 83.8% on TB 3.0); difficulty is calibrated against the latest frontier models
• Contributing institutions include Stanford, MIT, Princeton, Tsinghua, Peking, Oxford and Cambridge, led by the Terminal-Bench and Harbor Framework teams, with acknowledged reviewers, science advisors and AI research advisors from major AI labs
• Positioning: define the next generation of science coding agents — a "Claude Code for science" — and fill a measurement gap for research capability

Why it matters

• Accelerating scientific workflows is seen as the biggest lever for human progress, yet no benchmark specifically measured it; TBS provides a trackable standard
• Calibrated difficulty plus strong discrimination makes it effective for measuring future progress on scientific capability
• Global open review and a 920-to-70 selection ensure quality and resistance to saturation

2026-08-28 ~ 2026-08-29 · 19 related posts

• Stanford releases Terminal-Bench-Science: AI agent benchmark for research workflows — BenBlaiszik · 2026-08-28
• [source] TerminalBench-Science released: Opus 5 achieves 30% pass rate — BenBlaiszik · 2026-08-28
• TerminalBench-Science selection: 70 tasks chosen from 920 proposals — BenBlaiszik · 2026-08-28
• [source] TerminalBench-Science sets high bar, slashing model scores by 10+ points — BenBlaiszik · 2026-08-28
• [source] TerminalBench-Science v0.1: Opus 5 leads overall benchmarks — BenBlaiszik · 2026-08-28
• TerminalBench-Science thanks extensive reviewer and advisor team — BenBlaiszik · 2026-08-28
• Terminal-Bench-Science Aims to Fill the Measurement Gap in Science — BenBlaiszik · 2026-08-28
• Stanford Releases Terminal-Bench-Science: Claude Opus 5 Solves Only ~30% — BenBlaiszik · 2026-08-28
• Terminal-Bench-Science 0.1 releases 70 cross-domain science tasks — sanmikoyejo · 2026-08-28
• 70 tasks selected from 920 proposals for benchmark — sanmikoyejo · 2026-08-28
• Terminal-Bench-Science lowers model pass rates by over 10% — sanmikoyejo · 2026-08-28
• Terminal-Bench-Science built with global top university scientists — sanmikoyejo · 2026-08-28
• Terminal-Bench-Science launches: Claude Opus 5 solves only ~30% of research tasks — BenBlaiszik · 2026-08-28
• Terminal-Bench-Science introduced for scientific terminal benchmarking — BenBlaiszik · 2026-08-28
• Inside Terminal-Bench-Science: Defining the Next Gen of Science Agents — ajratner · 2026-08-28
• Stanford launches Terminal-Bench-Science: Claude Opus 5 solves only ~30% — ajratner · 2026-08-28
• Stanford's Terminal-Bench-Science: Claude Opus 5 solves only ~30% of research tasks — realsohamparekh · 2026-08-28

2 near-duplicate retellings: BenBlaiszik · WeijiaShi2