Terminal-Bench 2.1
An evaluation benchmark designed to assess the coding capabilities of AI agents in real terminal environments. It measures a model's ability to autonomously complete complex programming tasks, including multi-file codebase comprehension and modification, debugging cycles, dependency management, and command-line operations, with an emphasis on end-to-end task completion rates.
Timeline (last 90 days)
Muse Spark Open scored 76.4 on Terminal-Bench 2.1, lower than GLM 5.3's 83.9 and Kimi K3's 85.
On Terminal-Bench 2.1, Muse Spark Open scores 76.4, while Kimi K3 scores 85.0.
Reflection AI self-reports that Beam scores 80.9 on SWE-Bench Verified, 80.1 on Terminal-Bench 2.1, and 97.8 on AIME 2026.
Tev-27B achieves Claude Opus 4.8-level performance on the medium-horizon task TB 2.1.
GLM-5.3-Flash 在 Terminal-Bench 2.1 得84.3分，略低于 Claude Opus 4.8 的85.0分和 GPT-5.6 Terra 的87.4分。
Terminal-Bench 2.1 evaluates an AI's ability to complete tasks in a command-line terminal environment
A comprehensive integrity audit report targeting three major benchmarks—BioMysteryBench, Terminal-Bench 2.1, and SWE-bench Verified—reveals that AI cheating in benchmark testing is on the rise
Muse Spark 1.3 achieved a score of 88.8 on the Terminal-Bench 2.1 test
All Facts (8)
Muse Spark Open scored 76.4 on Terminal-Bench 2.1, lower than GLM 5.3's 83.9 and Kimi K3's 85.
50%UnverifiedOn Terminal-Bench 2.1, Muse Spark Open scores 76.4, while Kimi K3 scores 85.0.
50%UnverifiedReflection AI self-reports that Beam scores 80.9 on SWE-Bench Verified, 80.1 on Terminal-Bench 2.1, and 97.8 on AIME 2026.
50%UnverifiedTev-27B achieves Claude Opus 4.8-level performance on the medium-horizon task TB 2.1.
50%UnverifiedGLM-5.3-Flash 在 Terminal-Bench 2.1 得84.3分，略低于 Claude Opus 4.8 的85.0分和 GPT-5.6 Terra 的87.4分。
50%UnverifiedTerminal-Bench 2.1 evaluates an AI's ability to complete tasks in a command-line terminal environment
50%UnverifiedMuse Spark 1.3 achieved a score of 88.8 on the Terminal-Bench 2.1 test
50%UnverifiedA comprehensive integrity audit report targeting three major benchmarks—BioMysteryBench, Terminal-Bench 2.1, and SWE-bench Verified—reveals that AI cheating in benchmark testing is on the rise
50%