Benchmark · Agent tasks

Terminal-Bench

An agent harness running a model in a sandboxed shell, five trials or more on every task, until the task's tests pass. A row is one harness and model pair at one reasoning effort; the model is what's ranked.

Score
% of trials passed, higher is better
Run by
Stanford and the Laude Institute (the Terminal-Bench maintainers)
Status
Current, newest result 20 November 2026
Updated
as models are released
Data
the published file, Apache-2.0, read 20 November 2026
Method
www.tbench.ai/

Results: 4.0

26 models we list, each at its best setting, best first.

#ModelScoreRangeWho ran itSettingDate
1Claude Opus 5.5 A64.8%61.7% to 68.0%Independenteffort max, harness Claude Code, harness version 2.1.284, trials 3302026-09-30source
2Claude Sonnet 5.5 A61.8%58.9% to 64.8%Independenteffort max, harness Claude Code, harness version 2.1.284, trials 3302026-09-30source
3Kimi K3.1 Pro D60.9%56.8% to 65.0%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-10-30source
4GPT-6.1 Sol AA58.2%55.1% to 61.3%Independenteffort max, harness Codex, harness version 0.159.0, trials 3302026-09-30source
5Claude Fable 5.1 A57.9%54.1% to 61.6%Independenteffort max, harness Claude Code, harness version 2.1.257, trials 3302026-09-02source
6Gemini 4 Argon B54.9%51.9% to 57.8%Independenteffort high, harness Antigravity CLI, harness version 1.5.2, trials 3302026-11-13source
7GPT-6.1 Terra AA51.8%49.0% to 54.6%Independenteffort max, harness Codex, harness version 0.164.0, trials 3302026-10-28source
8GPT-6 Sol A49.4%46.2% to 52.6%Independenteffort max, harness Codex, harness version 0.159.0, trials 3302026-09-30source
9Muse Wonder D47.0%43.2% to 50.8%Independenteffort xhigh, harness Muse Code, harness version 1.4.14, trials 3302026-11-05source
10GLM-5.3 B41.8%38.6% to 45.0%Independenteffort max, harness Claude Code, harness version 2.1.207, trials 3302026-08-26source
11Claude Haiku 5.5 A40.0%36.3% to 43.7%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-06source
11DeepSeek V4.1 Flash C40.0%36.3% to 43.7%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-10-31source
13Tencent Hy4 B39.1%35.6% to 42.6%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-10source
14GLM-5.3-Flash B36.7%33.0% to 40.3%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-02source
15Mistral Large 4 C27.9%24.0% to 31.8%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-04source
16Kimi K3 B22.4%18.6% to 26.2%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-10source
17Gemini 3.8 Flash AA19.1%15.7% to 22.4%Independenteffort high, harness mini-SWE-agent, harness version 2.4.6, trials 3302026-09-01source
18GPT-6 Luna A16.4%13.6% to 19.1%Independenteffort max, harness Codex, harness version 0.159.0, trials 3302026-09-30source
19Muse Spark 1.3 D14.6%11.6% to 17.5%Independenteffort xhigh, harness Muse Code, harness version 1.4.1, trials 3302026-09-30source
20MiniMax M3 C7.3%4.6% to 10.0%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-10source
21DeepSeek V4 Pro D6.7%4.2% to 9.2%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-10source
22Beam 1.0 C6.4%4.7% to 8.0%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-09source
23Gemini 3.1 Pro (preview) AA2.7%1.1% to 4.4%Independenteffort high, harness Antigravity CLI, harness version 1.5.2, trials 3302026-11-14source
24Inkling-Small B0.9%0.0% to 1.9%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-06source
25Nemotron 3.5 Super VL C0.3%-0.3% to 0.9%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-09source
26Inkling B0.0%0.0% to 0.0%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-06source
Every setting, 27 rows
ModelAs the leaderboard names itScoreWho ran itSettingDate
Claude Opus 5.5anthropic/claude-opus-5-564.8%Independenteffort max, harness Claude Code, harness version 2.1.284, trials 3302026-09-30source
Claude Sonnet 5.5anthropic/claude-sonnet-5-561.8%Independenteffort max, harness Claude Code, harness version 2.1.284, trials 3302026-09-30source
Kimi K3.1 ProKimi K3.1 Pro60.9%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-10-30source
GPT-6.1 SolGPT-6.1 Sol58.2%Independenteffort max, harness Codex, harness version 0.159.0, trials 3302026-09-30source
Claude Fable 5.1anthropic/claude-fable-5-157.9%Independenteffort max, harness Claude Code, harness version 2.1.257, trials 3302026-09-02source
Gemini 4 ArgonGemini 4 Argon54.9%Independenteffort high, harness Antigravity CLI, harness version 1.5.2, trials 3302026-11-13source
GPT-6.1 TerraGPT-6.1 Terra51.8%Independenteffort max, harness Codex, harness version 0.164.0, trials 3302026-10-28source
GPT-6 SolGPT-6 Sol49.4%Independenteffort max, harness Codex, harness version 0.159.0, trials 3302026-09-30source
Muse WonderMuse Wonder47.0%Independenteffort xhigh, harness Muse Code, harness version 1.4.14, trials 3302026-11-05source
GLM-5.3GLM-5.341.8%Independenteffort max, harness Claude Code, harness version 2.1.207, trials 3302026-08-26source
Claude Haiku 5.5anthropic/claude-haiku-5-540.0%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-06source
DeepSeek V4.1 FlashDeepSeek V4.1 Flash 101740.0%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-10-31source
Tencent Hy4Hy439.1%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-10source
GLM-5.3-FlashGLM-5.3-Flash36.7%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-02source
GLM-5.3-FlashGLM-5.3-Flash35.8%Independenteffort none, harness Claude Code, harness version 2.1.285, trials 3302026-10-06source
Mistral Large 4Mistral Large 427.9%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-04source
Kimi K3Kimi K322.4%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-10source
Gemini 3.8 FlashGemini 3.8 Flash19.1%Independenteffort high, harness mini-SWE-agent, harness version 2.4.6, trials 3302026-09-01source
GPT-6 LunaGPT-6 Luna16.4%Independenteffort max, harness Codex, harness version 0.159.0, trials 3302026-09-30source
Muse Spark 1.3Muse Spark 1.314.6%Independenteffort xhigh, harness Muse Code, harness version 1.4.1, trials 3302026-09-30source
MiniMax M3MiniMax M37.3%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-10source
DeepSeek V4 ProDeepSeek V4 Pro 08136.7%Independenteffort max, harness Claude Code, harness version 2.1.311, trials 3302026-11-10source
Beam 1.0Beam 1.06.4%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-09source
Gemini 3.1 Pro (preview)gemini/gemini-3.1-pro-preview2.7%Independenteffort high, harness Antigravity CLI, harness version 1.5.2, trials 3302026-11-14source
Inkling-SmallInkling Small0.9%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-06source
Nemotron 3.5 Super VLNemotron 3.5 Super VL0.3%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-09source
InklingInkling0.0%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, trials 3302026-11-06source

Results: 4.1

24 models we list, each at its best setting, best first.

#ModelScoreRangeWho ran itSettingDate
1Claude Fable 5.2 A65.8%62.9% to 68.6%Independenteffort max, harness Claude Code, harness version 2.1.321, trials 3302026-11-10source
2Gemini 4 Carbon B64.5%61.3% to 67.8%Independenteffort high, harness Antigravity CLI, harness version 1.5.7, trials 3302026-11-14source
3GPT-6.2 Astra AA62.1%59.8% to 64.4%Independenteffort max, harness Codex, harness version 0.169.0, trials 3302026-11-17source
4GPT-6.2 Sol AA61.5%58.8% to 64.3%Independenteffort max, harness Codex, harness version 0.169.0, trials 3302026-11-17source
5Gemini 4 Neon B61.2%58.4% to 64.1%Independenteffort xhigh, harness Antigravity CLI, harness version 1.5.12, trials 3302026-11-20source
6Kimi K3.1 Pro D59.7%55.5% to 63.9%Independenteffort max, harness Claude Code, harness version 2.1.311, regrade yes, trials 3302026-11-12source
7Gemini 4 Hydrogen B56.4%53.4% to 59.3%Independenteffort high, harness Antigravity CLI, harness version 1.5.7, trials 3302026-11-13source
8Gemini 4 Argon B55.5%52.6% to 58.3%Independenteffort high, harness Antigravity CLI, harness version 1.5.2, regrade yes, trials 3302026-11-13source
9GPT-6.1 Terra AA53.3%50.2% to 56.4%Independenteffort max, harness Codex, harness version 0.164.0, regrade yes, trials 3302026-11-12source
10Muse Wonder D47.0%43.2% to 50.8%Independenteffort xhigh, harness Muse Code, harness version 1.4.14, regrade yes, trials 3302026-11-13source
11Claude Haiku 5.5 A40.0%36.4% to 43.6%Independenteffort max, harness Claude Code, harness version 2.1.311, regrade yes, trials 3302026-11-12source
12DeepSeek V4.1 Flash C38.5%34.7% to 42.2%Independenteffort max, harness Claude Code, harness version 2.1.311, regrade yes, trials 3302026-11-12source
13Tencent Hy4 B37.9%34.4% to 41.4%Independenteffort max, harness Claude Code, harness version 2.1.311, regrade yes, trials 3302026-11-12source
14GLM-5.3-Flash B37.0%33.3% to 40.6%Independenteffort max, harness Claude Code, harness version 2.1.311, regrade yes, trials 3302026-11-12source
15Muse Ember D30.6%26.8% to 34.4%Independenteffort xhigh, harness Muse Code, harness version 1.4.14, trials 3302026-11-10source
16Mistral Large 4 C27.6%23.7% to 31.4%Independenteffort max, harness Claude Code, harness version 2.1.311, regrade yes, trials 3302026-11-12source
17Kimi K3 B23.0%19.3% to 26.8%Independenteffort max, harness Claude Code, harness version 2.1.311, regrade yes, trials 3302026-11-12source
18MiniMax M3 C7.6%4.8% to 10.3%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, regrade yes, trials 3302026-11-12source
19DeepSeek V4 Pro D6.7%4.2% to 9.2%Independenteffort max, harness Claude Code, harness version 2.1.311, regrade yes, trials 3302026-11-12source
20Beam 1.0 C6.4%4.7% to 8.0%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, regrade yes, trials 3302026-11-12source
21Gemini 3.1 Pro (preview) AA2.7%1.1% to 4.4%Independenteffort high, harness Antigravity CLI, harness version 1.5.2, regrade yes, trials 3302026-11-15source
22Inkling-Small B0.9%0.0% to 1.9%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, regrade yes, trials 3302026-11-12source
23Nemotron 3.5 Super VL C0.3%-0.3% to 0.9%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, regrade yes, trials 3302026-11-12source
24Inkling B0.0%0.0% to 0.0%Independenteffort none, harness mini-swe-agent, harness version 2.4.6, regrade yes, trials 3302026-11-12source

A rank is among the models we list with a result, not a place on the whole leaderboard, and a model offered by several hosts is shown once, on its own listing. These results don't move our grades. Data: Terminal-Bench, Stanford and the Laude Institute, Apache 2.0, Apache-2.0.

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.