The honest view from inside technical leadership
LLM Leaderboard · Agentic coding

Terminal-Bench 4.1 leaderboard

Whether a model, driving an agent harness, can finish real multi-step tasks in a Linux terminal: building software, fixing environments, running and debugging code. The score is the share of tasks resolved.

← All models

Rankings

Snapshot 22 November 2026 · Scores copied from each benchmark's own leaderboard · How the CTO Index works

Score and price

#ModelProviderTerminal-BenchPrice $/M
1Copilot FusionGitHub71.2%Not listed
2Claude Fable 5.2Anthropic65.8%$8 / $40
3Claude Opus 5.5Anthropic65.8%$4 / $20
4Gemini 4 CarbonGoogle64.5%$2 / $10
5Claude Sonnet 5.5Anthropic63.9%$2 / $10
6GPT-6.2 AstraOpenAI62.1%$8 / $40
7GPT-6.2 SolOpenAI61.5%$2 / $10
8Gemini 4 NeonGoogle61.2%$1.50 / $7.50
9GPT-6.1 SolOpenAI60.9%$2 / $10
10GPT-6 AstraOpenAI60.3%$10 / $50
11Kimi K3.1 ProMoonshot AI59.7%$2 / $10
12Claude Fable 5.1Anthropic57.6%$10 / $50
13Gemini 4 HydrogenGoogle56.4%$1.50 / $7.50
14Gemini 4 ArgonGoogle55.5%$2 / $10
15Claude Opus 5Anthropic53.3%$5 / $25
16GPT-6.1 TerraOpenAI53.3%$1 / $5
17GPT-6 SolOpenAI50.3%$2 / $10
18Muse WonderMeta47.0%$4 / $20
19Claude Fable 5Anthropic44.8%$10 / $50
20GLM-5.3Z.ai41.2%$1.40 / $4.40
21Claude Haiku 5.5Anthropic40.0%$0.10 / $0.50
22Grok 4.7xAI38.8%$2 / $6
23GPT-5.6 SolOpenAI38.5%$4 / $20
24DeepSeek V4.1 FlashDeepSeek38.5%$0.15 / $0.60
25Hy4Tencent37.9%Not listed
26GLM-5.3-FlashZ.ai37.0%$0.02 / $0.07
27Muse EmberMeta30.6%$0.10 / $0.50
28Mistral Large 4Mistral27.6%$0.70 / $2.10
29Qwen3.8-MaxAlibaba27.0%$2 / $6
30MiMo-V2.6-ProXiaomi27.0%$0.43 / $0.87
31Claude Opus 4.8Anthropic23.6%$5 / $25
32Kimi K3Moonshot AI23.0%$3 / $15
33Grok 4.6xAI19.7%$1.60 / $4.80
34Gemini 3.8 FlashGoogle18.2%$0.75 / $3.75
35MiMo-V2.6-FlashXiaomi16.1%$0.14 / $0.28
36GPT-6 LunaOpenAI15.2%$0.10 / $0.50
37Muse Spark 1.3Meta14.8%$1.25 / $4.25
38Inkling 2 SmallThinking Machines13.0%$0.30 / $1
39Gemini 3.7 FlashGoogle12.4%$0.75 / $3.75
40Claude Sonnet 5Anthropic12.1%$2 / $10
41Beam 1.0Reflection AI6.4%$0.30 / $1

Other benchmarks

Frequently asked questions

What does Terminal-Bench 4.1 measure?

Whether a model, driving an agent harness, can finish real multi-step tasks in a Linux terminal: building software, fixing environments, running and debugging code. The score is the share of tasks resolved.

Which model has the highest Terminal-Bench 4.1 score?

Copilot Fusion from GitHub leads Terminal-Bench 4.1 at 71.2%, ahead of Claude Fable 5.2 (65.8%).

Why does Terminal-Bench 4.1 matter to engineering leaders?

This is the closest public proxy for how a coding agent behaves inside your repository and CI. It rewards finishing the job, not writing a plausible-looking snippet.

What are the limitations of Terminal-Bench 4.1?

Scores depend on the agent harness (Claude Code, Codex, Antigravity CLI, mini-SWE-agent and others), so part of the gap between models is the harness. Error bars are roughly ±3 points; treat differences smaller than that as ties.