Benchmark leaderboard - 23 entries
Terminal-Bench leaderboard
Terminal-Bench is the benchmark for testing AI agents in real terminal environments, evaluating how well agents can handle real-world, end-to-end tasks (from compiling code to training models and setting up servers) autonomously. As of October 17, 2026, Opus 5.5 ranks first on the Terminal-Bench 4.0 board with 64.8% ± 3.1%.
Last verified from tbench.ai. Read the caveats
Terminal-Bench 4.0
23 entries, in the board's order
| 1 | Opus 5.5 Rank 1; agent: Claude Code; effort: max; tokens 8.0B; cost $4.7k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 64.8% ± 3.1% |
| 2 | Sonnet 5.5 Rank 2; agent: Claude Code; effort: max; tokens 19.4B; cost $7.3k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 61.8% ± 2.9% |
| 3 | GPT-6 Astra Rank 3 (tie); agent: Codex; effort: max; tokens 1.5B; cost $3.3k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 58.2% ± 2.8% |
| 4 | GPT-6.1 Sol Rank 3 (tie); agent: Codex; effort: max; tokens 1.5B; cost $0.6k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 58.2% ± 3.1% |
| 5 | Fable 5.1 Rank 5; agent: Claude Code; effort: max; tokens 2.7B; cost $6.2k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 57.9% ± 3.8% |
| 6 | Opus 5 Rank 6; agent: Claude Code; effort: xhigh; tokens 6.9B; cost $6.1k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 53.9% ± 3.2% |
| 7 | GPT-6 Sol Rank 7; agent: Codex; effort: max; tokens 5.0B; cost $1.5k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 49.4% ± 3.2% |
| 8 | Fable 5 Rank 8; agent: Claude Code; effort: max; tokens 3.8B; cost $7.3k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 44.5% ± 3.8% |
| 9 | GLM-5.3 Rank 9; agent: Claude Code; effort: max; tokens 8.7B; cost $2.7k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 41.8% ± 3.2% |
| 10 | Grok 4.7 Rank 10; agent: Grok Build; effort: xhigh; tokens 5.5B; cost $3.7k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 37.6% ± 3.5% |
| 11 | GPT-5.6 Sol Rank 11; agent: Codex; effort: max; tokens 4.4B; cost $2.5k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 37.3% ± 3.8% |
| 12 | GLM-5.3-Flash Rank 12; agent: Claude Code; effort: none; tokens 2.9B; cost $2.9k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 35.8% ± 3.5% |
| 13 | Qwen3.8-Max-0902 Rank 13; agent: Claude Code; effort: max; tokens 3.0B; cost $4.6k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 27.0% ± 3.8% |
| 14 | Opus 4.8 Rank 14; agent: Claude Code; effort: max; tokens 6.4B; cost $6.5k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 23.6% ± 3.6% |
| 15 | GPT-5.6 Terra Rank 15; agent: Codex; effort: max; tokens 5.7B; cost $1.7k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 21.5% ± 3.3% |
| 16 | Grok 4.6 Rank 16; agent: Grok Build; effort: high; tokens 4.0B; cost $3.6k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 20.3% ± 3.1% |
| 17 | Gemini 3.8 Flash Rank 17; agent: mini-SWE-agent; effort: high; tokens 17.2B; cost $1.8k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 19.1% ± 3.4% |
| 18 | GPT-5.6 Luna Rank 18; agent: Codex; effort: max; tokens 11.6B; cost $0.3k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 17.3% ± 2.8% |
| 19 | GPT-6 Luna Rank 19; agent: Codex; effort: max; tokens 3.9B; cost $0.1k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 16.4% ± 2.7% |
| 20 | Muse Spark 1.3 Rank 20; agent: Muse Code; effort: xhigh; tokens 9.0B; cost not shown; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 14.5% ± 2.9% |
| 21 | Grok 4.5 Rank 21 (tie); agent: Grok Build; effort: high; tokens 3.4B; cost $2.1k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 12.4% ± 2.6% |
| 22 | Sonnet 5 Rank 21 (tie); agent: Claude Code; effort: max; tokens 21.6B; cost $9.6k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 12.4% ± 3.1% |
| 23 | Gemini 3.7 Flash Rank 23; agent: mini-SWE-agent; effort: high; tokens 11.1B; cost $1.3k; date = model release date shown on leaderboard; read 2026-10-17 Official leaderboard | 11.2% ± 2.4% |
About this board
- Ranks 24-25: the official Terminal-Bench 4.0 leaderboard (https://www.tbench.ai/leaderboard, read 2026-10-17) lists only 23 entries, so there is no top 25.
- Submission or evaluation dates are not shown on the leaderboard. The date column shows the model release date instead.
- The description comes from the official GitHub README (https://github.com/laude-institute/terminal-bench), because tbench.ai shows only the tagline 'A benchmark to measure and evolve with the frontier of agent work' and https://www.tbench.ai/about returns 404.
FAQ
What does Terminal-Bench measure?
Terminal-Bench is the benchmark for testing AI agents in real terminal environments, evaluating how well agents can handle real-world, end-to-end tasks (from compiling code to training models and setting up servers) autonomously.
Which model leads Terminal-Bench?
As of October 17, 2026, Opus 5.5 ranks first on the Terminal-Bench 4.0 board with 64.8% ± 3.1%. That is the board's own order, not a Developers Digest ranking.
Where do these Terminal-Bench scores come from?
All 23 entries come from the official leaderboard at tbench.ai. Every entry links to its source, which was read on October 17, 2026, and lists the setting it was run under.
Related reading
- DeepSeek V4.1-Flash-1017: Benchmarks vs Opus 5, Pricing, Switch?DeepSeek V4.1-Flash-1017 jumps 11 points on Terminal-Bench 4.0 at the same deepseek-flash price. Benchmarks vs Opus 5, K3, GLM-5.3 and when to switch.
- Terminal-Bench Shows Harness Scaling Is the Coding-Agent Benchmark NowTerminal-Bench and StateM show harness scaling now drives coding-agent scores. Why runbooks, durable state and recovery loops matter as much as model choice.
- GPT-5.6 Sol Ultra Coming to Codex with Cooperative SubagentsGPT-5.6 Sol Ultra is coming to Codex with trained subagents that communicate during tasks, reportedly hitting 91.9% on Terminal-Bench 2.1. What HN thinks.
- Muse Spark Open Weights: 816B Model, Benchmarks, and How to Run ItMeta released Muse Spark Open weights on October 15, 2026: 816B total, 42B active, 1M context. Run recipes, hardware floor, license catch, and where it trails.
- Claude Code Subagent Model: Put Explore on Haiku 5.5Claude Code's Explore subagent runs on your main model, and CLAUDE_CODE_SUBAGENT_MODEL alone won't move it. Three ways to put it on Haiku 5.5, with test runs.
- Cheapest Subagent Model: Haiku 5.5 or GPT-6 Luna? Routing, Compaction and Cost per TaskWhich cheap model belongs under your orchestrator? Haiku 5.5 vs GPT-6 Luna as a subagent or worker: routing by prompt length, compaction, tool-call gotchas and cost per agent task.
Sources
Last verified . Each row links to the page its score was copied from.
Newsletter
Get new benchmark results as they land
One email per week. New model releases, the leaderboard moves that matter and what they mean for what you build.