Stanford University researchers, working with the team behind Terminal-Bench, published Terminal-Bench-Science 0.1, an evaluation that scores AI agents against 70 research workflows pulled from scientists’ own projects rather than textbook exercises. The top performer, Claude Opus 5 running inside Claude Code, resolved 30.0 percent of the tasks. Every other system evaluated finished below that, and the gap between this benchmark and existing coding evals is itself part of the story.

This is a 0.1 release, the first of what Terminal-Bench-Science describes as a continuous benchmark rather than a one-time paper. The builders say tasks were deliberately calibrated to challenge current frontier models, and the results back that framing: no system cleared a third of the workload, and most cleared far less.

The full leaderboard, using each model’s best-performing agent scaffold:

Terminal-Bench-Science says its ranking spreads systems apart at roughly the same rate its software-engineering sibling, Terminal-Bench 3.0, does, even though every model tested on both saw its score fall by over 10 points here. The announcement frames that drop as deliberate calibration, not noise: reviewers built the task set to sit past what today’s agents solve easily.

Cost and token use diverge from raw accuracy. Claude Opus 5’s full 70-task run cost roughly $7,000, the highest of any system evaluated. GPT-5.6 Sol matched Claude Fable 5’s resolution rate at under a third of the price, $4,200 against $14,200. Claude Fable 5, in turn, matched GPT-5.6 Sol’s score while burning through 6.4 billion tokens against Sol’s 8.4 billion. Only Kimi K3 and Opus 5 showed up as efficient on both the cost curve and the token curve. No other system managed it on either measure.

Results also break out by scientific domain. Systems from Anthropic and OpenAI claimed the two best rankings in four of the five domains. The exception is engineering, where Grok 4.6 pulled even with GPT-5.6 Sol for the runner-up slot at 14.8 percent while spending less compute and fewer tokens to get there. Claude Opus 5 led everywhere except the math-focused category, which Claude Fable 5 (33.3 percent) and GPT-5.6 Sol (31.4 percent) took instead.

The task construction process explains why the benchmark is small relative to its ambitions. Contributors submitted 920 task proposals through an open GitHub process; reviewers approved 464 for implementation, 386 were turned into pull requests, and only 70 survived domain review, technical review, and a final bar-raiser check to reach the 0.1 release. Terminal-Bench-Science attributes that attrition to the difficulty of writing tasks that are scientifically meaningful, hard for frontier agents, and precisely verifiable at the same time.

That construction method is the real news underneath the leaderboard. Static benchmarks built from published problem sets are vulnerable to memorization and to labs quietly optimizing toward the test once scores become a marketing asset. Sourcing tasks directly from working scientists, then retiring them once agents saturate them, is a structural attempt to keep that from happening here. This issue of AI Insiders also covers a Google DeepMind report on double-blind evaluation, and taken together, the two pieces point to the same trend: measuring agent capability honestly is becoming its own research problem, not a solved formality that labs handle in a footnote.

Terminal-Bench-Science is now taking submissions for its 0.2 release, with October 5, 2026 set as the cutoff for new pull requests, and says each future round will be calibrated against whichever models are strongest at the time. Teams building research-automation agents should treat the 30 percent ceiling as the current state of the art, not a target already cleared, and should watch the domain-by-domain breakdown rather than the headline number when picking a model for a specific field.

Terminal-Bench-Science published these findings in its Terminal-Bench-Science 0.1 announcement.