ai.MODEST PRAISE. PUBLIC NUMBERS.

← Back to the scoreboard

Terminal-Bench 2.0

Data as of 2026-10-14 / manually checked

4 models · 7 benchmarks · zero composite scores

pretty greatpretty okaynot yetTap a verdict for the number and source.

Completes difficult tasks using a terminal.

Why it matters

Tests the work between writing code and getting it to run.

How to read it

Percentage of tasks resolved. Agent harness matters. Higher is better.

Run by Stanford / Laude Institute. Last checked 2026-10-14.

Visit the benchmark source ↗

MODEL / VENDORTerminal-Bench 2.0terminal work

OpenAI

GPT-5.2 Thinking

Pretty great at coding, abstract reasoning, science, math, long context.

Terminal-Bench 2.0 not reported

OpenAI

GPT-5.1 Thinking

Pretty great at coding, science.

Terminal-Bench 2.0 not reported

Anthropic

Claude Opus 4.6

Pretty great at coding, terminal work.

Terminal-Bench 2.0

pretty great

65.4%

Vendor-reported. Terminus-2 harness, adaptive thinking at maximum effort, 1 million tokens per task; 1× guaranteed / 3× ceiling resources, five attempts per task.

Source ↗Observed 2026-10-14

Google

Gemini 4 Argon High

Pretty great at conversation.

Terminal-Bench 2.0 not reported

The top score means the top sourced score in this snapshot. Different harnesses, budgets and samples can change results. A blank cell is a blank cell.