Atlas

Benchmarks

Terminal-Bench 4.1

Tbench

Terminal-Bench 4.1 is a minor revision of Terminal-Bench 4.0 that leaves every agent-facing task environment unchanged and hardens the verifiers against reward hacking: Harbor strips dangerous symlinks and permissions from agent artifacts, submitted code runs as an untrusted user in a separate process before a trusted grader scores it, and several verifier flakes are fixed. The maintainers regraded the stored 4.0 trials, changing 162 of 11,550, and every agent's 4.1 score fell within its 4.0 95% confidence interval; trials without a regrade keep their 4.0 score. Atlas files it as a component of the Terminal-Bench 4.0 family, which keeps 4.0 as its primary measurement. Scores are mean resolution rates across repeated trials.

Models
1
Top score
57.0
Leader
Gemini 4 Hydrogen
Updated
Nov 12, 2026
Resolution
—
Saturation
—
Signal
94thpercentile
192 benchmarks from 0.8 to 3.5218 benchmarks from 3.5 to 6.1128 benchmarks from 6.1 to 8.895 benchmarks from 8.8 to 11.567 benchmarks from 11.5 to 14.128 benchmarks from 14.1 to 16.814 benchmarks from 16.8 to 19.49 benchmarks from 19.4 to 22.111 benchmarks from 22.1 to 24.72 benchmarks from 24.7 to 27.46 benchmarks from 27.4 to 30.13 benchmarks from 30.1 to 32.73 benchmarks from 32.7 to 35.45 benchmarks from 35.4 to 38.01 benchmark from 38.0 to 40.71 benchmark from 40.7 to 43.41 benchmark from 43.4 to 46.01 benchmark from 62.0 to 64.6

50th strongest of 785

Signal-to-noise +18.29 for the Terminal-Bench 4.0 family

Difficulty
Not estimated

Another benchmark represents this family in the fit.

Organizations

Not enough dated models to plot (1 of 1 have a release date).

#ModelSettingScoreSource
1Gemini 4 Hydrogenhigh · Mini-SWE-Agent vision harness
57.0
Google Gemini 4 Hydrogen evaluation: gemini-4-hydrogen
Loading Atlas data…