activeAgentsScience

Terminal-Bench-Science 0.1

Metric: Accuracy (higher is better)Introduced: 2026

Stanford-led community benchmark of 70 tasks drawn from scientific research workflows across the life, physical, earth, mathematical, and engineering sciences. An agent CLI works from natural-language instructions in a self-contained environment; artifacts are graded against hidden task-specific tests. Distinct from Terminal-Bench 2.x/3.0/4.0.

Models ranked

2

tracked on this benchmark

Score band

75.3 – 74.6

best → lowest tracked

Snapshot trend

—

need ≥2 snapshots

Leaderboard

Tracked models ranked by Accuracy (higher is better).

Compare candidates
#Model variant and provenanceScore
1
Claude Mythos 5.2

Configuration: claude-mythos-5-2 (Claude Code --bare, max thinking, no safeguards)

Version: Terminal-Bench-Science 0.1Harness: Claude Code --bare mode, no internetEvaluator: AnthropicObserved: Nov 9, 2026Confidence: self-reportedSource

Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.6 and launch table (Mythos 5.2 parenthetical): 75.3% without the Fable safeguards. Same 70-task, 10-trial setup. Separate from the Fable 5.2 74.6% row.

75.3
2
Claude Fable 5.2

Configuration: claude-fable-5-2 (Claude Code --bare, max thinking, safeguards on)

Version: Terminal-Bench-Science 0.1Harness: Claude Code --bare mode, no internetEvaluator: AnthropicObserved: Nov 9, 2026Confidence: self-reportedSource

Notes: Anthropic self-reported (lab-published). System card + launch table, retrieved 2026-11-09. Does not move /best rankings (self-reported / raon rule). System card §8.6 and launch table: 74.6% with safeguards on. 70 Stanford-led tasks; 10 trials per task (700 trials); clustered SE ±4.3 points. Claude Code --bare, max thinking, no-internet-access setup as in Terminal-Bench 4.0. Fallback policy affected 1.3% of requests (4.1% of trials).

74.6

How to read this benchmark

This benchmark scores models where higher is better. Use scores for directional filtering and shortlisting, not universal quality ranking; then validate pricing, context window, provider availability, and fit for your workload.

Trust this score when

  • There is a fresh timestamped snapshot (or multiple snapshots) for this benchmark.
  • The model list covers the same version family you can actually deploy today.
  • Top candidates overlap with your required routing and feature requirements.

Be cautious when

  • There is only one benchmark snapshot or the dataset appears stale.
  • The benchmark metric direction is opposite of your decision objective.
  • The score difference between options is narrow and likely within implementation variance.

Related benchmarks

AA Coding Agent Index
Artificial Analysis Coding Agent Index
Agents' Last Exam
Android Daily
AndroidWorld
AutomationBench
BFCL v3
Berkeley Function Calling Leaderboard v3

Last reviewed: Nov 9, 2026

Resources

Website