Terminal-Bench 2.1
GPT-6 Astra 87.4% leads since Sep 2026 In use
89 hard terminal tasks across 16 categories in containers. Hidden tests verify the final state. 5 trials per task.
Fixed 28 of the 89 tasks in 2.0. Scores moved by up to 12 points, so the two are kept apart.
In the authors’ words
AI agents may soon become capable of autonomously completing valuable, long-horizon tasks in diverse domains. Current benchmarks either do not measure real-world tasks, or are not sufficiently difficult to meaningfully measure frontier models. To this end, we present Terminal-Bench 2.0: a carefully curated hard benchmark composed of 89 tasks in computer terminal environments inspired by problems from real workflows. Each task features a unique environment, human-written solution, and comprehensive tests for verification. We show that frontier models and agents score less than 65% on the benchmark and conduct an error analysis to identify areas for model and agent improvement. We publish the dataset and evaluation harness to assist developers and researchers in future work at https://www.tbench.ai/ .
Made by Stanford + Laude Institute Released May 2026 Terminal tasks All tasks public Last in a launch post Oct 2026
Last model added Oct 2026
Scores refreshed 2 Nov 2026
- Level
- Expert Specialists with years of training, often over hours of work
- Scored
- Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
- Answers
- Agent task Many steps in a terminal, browser, desktop, app, game or robot.
13
Models scored on its own leaderboard
53
Launch posts list it, from 27 labs
22
System cards report it, from 11 labs
GPT-6 Astra leads with 87.4% among models released by Sep 2026
Best score per model from the benchmark's own leaderboard, read on 2 Nov 2026. Where a model ran with several settings or harnesses, we keep its best.
The newest model with a score here came out in Sep 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- Claude Opus 5.5
- GPT-6.1 Terra
- GPT-6.1 Sol
- Claude Sonnet 5.5
- and 11 more
- 1 GPT-6 Astra high 87.4%
- 2 Claude Fable 5 xhigh 83.8%
- 3 GPT-5.5 xhigh 83.2%
- 4 Grok 4.5 high 79.3%
- 5 Claude Opus 4.8 high 78.9%
- 6 GPT-5.6 Terra max 78.4%
- 7
Muse Spark 1.1 xhigh 76.2% - 8 GPT-5.6 Luna max 75.7%
- 9 Claude Sonnet 5 high 74.6%
- 10
Gemini 3 Pro high 73.9% - 11 Claude Opus 4.7 max 68.9%
- 12 Gemini 3.1 Pro high 65.8%
- 13 GLM-5.1 max 58.6%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-6 Astra | 87.4% | high, Codex | Sep 2026 | Official leaderboard |
| 2 | Claude Fable 5 | 83.8% | xhigh, Claude Code | Jun 2026 | Official leaderboard |
| 3 | GPT-5.5 | 83.2% | xhigh, Codex | Apr 2026 | Official leaderboard |
| 4 | Grok 4.5 | 79.3% | high, Cursor CLI | Jul 2026 | Official leaderboard |
| 5 | Claude Opus 4.8 | 78.9% | high, Claude Code | May 2026 | Official leaderboard |
| 6 | GPT-5.6 Terra | 78.4% | max, Codex | Jul 2026 | Official leaderboard |
| 7 | Muse Spark 1.1 | 76.2% | xhigh, mini-SWE-agent | Jul 2026 | Official leaderboard |
| 8 | GPT-5.6 Luna | 75.7% | max, Codex | Jul 2026 | Official leaderboard |
| 9 | Claude Sonnet 5 | 74.6% | high, Claude Code | Jun 2026 | Official leaderboard |
| 10 | Gemini 3 Pro | 73.9% | high, Terminus 2 | Nov 2025 | Official leaderboard |
| 11 | Claude Opus 4.7 | 68.9% | max, Claude Code | Apr 2026 | Official leaderboard |
| 12 | Gemini 3.1 Pro | 65.8% | high, Gemini CLI | Feb 2026 | Official leaderboard |
| 13 | GLM-5.1 | 58.6% | max, Claude Code | Apr 2026 | Official leaderboard |
Labs report scores for 53 models in their cards. We plot every source in the charts below.
3 models have held number one since its release in May 2026
Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.
OpenAI Anthropic Google DeepMind Cognition
The best score passed 85% in Jun 2026
All the scores we have, one dot each, so a model tested at several settings appears more than once.
On the benchmark’s own leaderboard Reported by the lab in its model card
Best so far
27 labs put it in a launch post in the last six months
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| 27 Oct 2026 | LFM3-1.7B | 12.9 | Terminal-Bench 2.1 (OpenCode) |
| 15 Oct 2026 | Muse Spark Open | 76.4 | Terminal-Bench 2.1 Terminus 2 agent |
| 15 Oct 2026 |  Nemotron 3.5 Bolt 74B-A9B | 51 | Terminal-Bench 2.1 |
| 15 Oct 2026 | Nemotron 3.5 Super VL | 60.7 | Terminal-Bench 2.1 (Terminus 2) |
| 13 Oct 2026 |  Ling-3.1-flash | 81.2 | Terminal-Bench-2.1 |
| 5 Oct 2026 | Beam 1.0 | 80.1 | Terminal Bench v2.1 |
| 22 Sep 2026 | MiMo-V2.6 Pro | 89.9 | Terminal Bench 2.1 |
| 20 Sep 2026 |  Step 5 Preview | 85 | Terminal-Bench v2.1 |
| 13 Sep 2026 |  Intern-S2-397B | 64 | TerminalBench 2.1 (Terminus 2) |
| 10 Sep 2026 | SWE-2 | 92.8 | Terminal-Bench 2.1 |
| 10 Sep 2026 |  DeepSeek V4.1 Flash | 90.6 | Terminal-Bench 2.1 (Pass@1) |
| 2 Sep 2026 | Muse Spark 1.3 | 88.8 | Terminal-Bench 2.1 (Agentic terminal coding) |
| 2 Sep 2026 | Gemini 3.8 Flash | 89.4 | Terminal-bench 2.1 |
| 28 Aug 2026 |  Tencent Hy4 preview | 85.4 | Terminal-Bench 2.1 |
| 26 Aug 2026 | GLM-5.3-Flash | 84.3 | Terminal Bench 2.1 (Claude Code 2.1.207) |
| 25 Aug 2026 | Granite 4.2 30B | 29.2 | Terminal-Bench 2.1 |
| 21 Aug 2026 | DeepSeek-V4-Flash-Vision-Exp | 83.9 | Terminal Bench 2.1 |
| 14 Aug 2026 |  Qwen 3.8 27B | 73 | Agentic terminal coding Terminal Bench 2.1 (Terminus) |
| 14 Aug 2026 | GLM-5.3 | 88.2 | Terminal Bench 2.1 |
| 13 Aug 2026 | DeepSeek V4 Pro 0813 | 87.9 | Terminal Bench 2.1 |
| 13 Aug 2026 | Gemini 3.7 Flash | 85.8 | Terminal-bench 2.1 |
| 11 Aug 2026 | Nemotron 3.5 Lightning 30B-A3B | 24.6 | Terminal-Bench 2.1 |
| 11 Aug 2026 |  Solar Pro 4 | 57 | Terminal-Bench v2.1 (Artificial Analysis score quoted by Upstage) |
| 10 Aug 2026 | Muse Glimmer | 51.7 | TerminalBench 2.1 (with terminus2) |
| 5 Aug 2026 | Muse Spark 1.2 | 82.9 | Terminal-Bench 2.1 (Muse Code harness) |
| 3 Aug 2026 | Qwen 3.8 Max | 86.6 | Terminal Bench 2.1 |
| 31 Jul 2026 |  K-EXAONE 2.0 | 43.8 | Terminal-Bench 2.1 |
| 31 Jul 2026 | DeepSeek V4 Flash 0731 | 82.7 | Terminal Bench 2.1 |
| 30 Jul 2026 |  Inkling-Small | 64.7 | Terminal Bench 2.1 (best harness) |
| 23 Jul 2026 |  KAT-Coder-V2.5-Dev | 41 | Terminal-Bench 2.1, mean of Terminus-2 32.60 / Claude Code 49.44 |
| 21 Jul 2026 |  Laguna S 2.1 | 70.2 | Terminal-Bench 2.1 |
| 21 Jul 2026 | Gemini 3.5 Flash-Lite | 54 | Terminal-bench 2.1 (Terminus-2 harness) |
| 21 Jul 2026 | Gemini 3.6 Flash | 78 | Terminal-bench 2.1 (Terminus-2 harness) |
| 16 Jul 2026 | Kimi K3 | 88.3 | Terminal-Bench 2.1 |
| 16 Jul 2026 | Grok 4.5 | 83.3 | Terminal Bench 2.1 |
| 15 Jul 2026 | Inkling | 63.8 | Terminal Bench 2.1 (best harness) |
| 9 Jul 2026 | GPT-5.6 Luna | 84.7 | Terminal-Bench 2.1 |
| 9 Jul 2026 | GPT-5.6 Terra | 87.4 | Terminal-Bench 2.1 |
| 9 Jul 2026 | Muse Spark 1.1 | 80 | Terminal-Bench 2.1 (Agentic terminal coding) |
| 9 Jul 2026 | GPT-5.6 Sol | 91.9 | Terminal-Bench 2.1 (Sol Ultra, 4 agents) |
| 8 Jul 2026 | SWE-1.7 | 81.5 | Terminal-Bench 2.1 (Devin CLI, timeout 4h) |
| 6 Jul 2026 | KAT-Coder-Pro V2.5 | 60.7 | Terminal-Bench 2.1 (AA Coding Index category) |
| 5 Jul 2026 |  LongCat-2.0 | 70.8 | Terminal-Bench 2.1 |
| 2 Jul 2026 | Tencent Hy3 | 71.7 | Terminal-Bench 2.1 |
| 30 Jun 2026 | Claude Sonnet 5 | 80.4 | Agentic coding Terminal-Bench 2.1 |
| 23 Jun 2026 |  Seed2.1 Pro | 71 | Terminal Bench 2.1 |
| 16 Jun 2026 | GLM-5.2 | 82.7 | Terminal Bench 2.1 (Best Reported Harness) |
| 9 Jun 2026 | Claude Fable 5 | 88 | Agentic coding Terminal-Bench 2.1 |
| 4 Jun 2026 | Nemotron 3 Ultra | 56.4 | Terminal Bench 2.1 |
| 1 Jun 2026 |  MiniMax-M3 | 66 | Terminal Bench 2.1 |
| 29 May 2026 | Step 3.7 Flash | 59.6 | Terminal-Bench 2.1 |
| 28 May 2026 | Claude Opus 4.8 | 74.6 | Agentic terminal coding Terminal-Bench 2.1 |
| 19 May 2026 | Gemini 3.5 Flash | 76.2 | Terminal-bench 2.1 (Terminus-2 harness) |
22 system cards report it, from 11 labs
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 19 Oct 2026 | Muse Glimmer technical report | 51.7 | TerminalBench 2.1 (Terminus2), high, 54.3 xhigh (Tables 2 and 5) | Capabilities |
| 15 Oct 2026 | Nemotron 3.5 Super VL technical report | 60.7 | Terminal-Bench 2.1, Terminus 2 harness, 3 hour timeout, mean of 8 runs (Table 10) | Capabilities |
| 15 Oct 2026 | Nemotron 3.5 Lightning 30B-A3B technical report | 24.4 | Terminal-Bench v2.1, BF16 acc. (Table 22) | Capabilities |
| 15 Oct 2026 | Nemotron 3.5 Bolt 74B-A9B technical report | 51 | Terminal-Bench 2.1, released 24GB checkpoint (Table 7). BF16 53.93 | Capabilities |
| 22 Sep 2026 | MiMo-V2.6 Pro technical report | 89.9 | Terminal Bench 2.1 (Table 3) | Capabilities |
| 10 Sep 2026 | DeepSeek V4.1 Flash technical report | 90.6 | Terminal-Bench 2.1 (Pass@1), DeepSeek Harness Minimal (Table 3, Table 4) | Capabilities |
| 2 Sep 2026 | Gemini 3.8 Flash model card | 89.4 | Terminal-bench 2.1 (results table, p.5) | Capabilities |
| 13 Aug 2026 | Gemini 3.7 Flash model card | 85.8 | Terminal-bench 2.1 (results table, p.5) | Capabilities |
| 5 Aug 2026 | K-EXAONE 2.0 technical report | 43.8 | Terminal-Bench 2.1 (Table 6, Reasoning mode) | Capabilities |
| 30 Jul 2026 | Inkling-Small model card | 64.7 | Terminal Bench 2.1, best harness (5. Evaluations table) | Capabilities |
| 27 Jul 2026 | Kimi K3 technical report | 88.3 | Terminal-Bench 2.1, best score across harnesses (Table 2) | Capabilities |
| 21 Jul 2026 | Gemini 3.6 Flash model card | 78 | Terminal-bench 2.1, Terminus-2 harness (results table, p.4) | Capabilities |
| 21 Jul 2026 | Gemini 3.5 Flash-Lite model card | 54 | Terminal-bench 2.1, Terminus-2 harness (results table, p.4) | Capabilities |
| 15 Jul 2026 | Inkling model card | 63.8 | Terminal Bench 2.1, best harness, contaminated rollouts scored 0 (5. Evaluations table, effort 0.99) | Capabilities |
| 14 Jul 2026 | Grok 4.5 model card | 83.3 | Terminal-Bench 2.1, task success rate, high, Grok Build harness (2.9) | Capabilities |
| 9 Jul 2026 | Muse Spark 1.1 system card | 80 | Terminal-Bench 2.1 (Figure 44) | Capabilities |
| 30 Jun 2026 | Claude Sonnet 5 system card | 80.4 | Terminal-Bench 2.1, mean reward over 5 attempts on 89 tasks at xhigh effort, mini-SWE-agent harness (Table 8.1.A, 8.3) | Capabilities |
| 23 Jun 2026 | Seed2.1 Pro model card | 71 | Terminal-Bench 2.1 (Table 5) | Capabilities |
| 9 Jun 2026 | Nemotron 3 Ultra technical report | 56.4 | Terminal Bench 2.1 (Table 10) | Capabilities |
| 9 Jun 2026 | Claude Fable 5 system card | 84.3 | Terminal-Bench 2.1, mini-SWE-agent harness, mean reward over 5 attempts on 89 tasks, high effort, 20.9% of trials hit a safety refusal and fell back to Opus 4.8 (Table 8.1.A, 8.3) | Capabilities |
| 28 May 2026 | Claude Opus 4.8 system card | 74.6 | Terminal-Bench 2.1 mean reward on Harbor with the Terminus-2 harness, high effort, 5 attempts on each of 89 tasks (Table 8.1.A, 8.3) | Capabilities |
| 19 May 2026 | Gemini 3.5 Flash model card | 76.2 | Terminal-bench 2.1, Terminus-2 harness (results table, p.4) | Capabilities |
Versions
We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.
Anthropic OpenAI Google DeepMind Cognition
More in Terminal agents
- SkillsBench In use 86-87 containerized expert tasks in 11 domains, run with and without Agent Skills. Verifier tests give pass rate and skill lift. GPT-5.5 67.3%
- Terminal-Bench Hard In use Agents complete 44 hard tasks from Terminal-Bench Core in the Terminus 2 harness. Artificial Analysis scores pass@1 averaged over 3 repeats. Gemma 4 31B IT 36%
- Harbor-Index In use Agents solve about 80 hard tasks drawn from 29 benchmarks run through Harbor adapters, such as HLE, GSO and SWE-bench Pro. Each task keeps its source verifier. GPT-5.5 28.1%
- CompileBench Not in model cards Agents build real open-source projects through dependency problems, legacy toolchains and complex build systems, for example compiling 22-year-old code. Claude Opus 4.1 100%
- Long-Horizon-Terminal-Bench 2.0 Not in model cards Agents work for up to 3 hours per task on 77 container tasks, including 30 tasks that rebuild a scientific tool from its outputs. Hidden verifiers give partial-credit rewards. GPT-6 Astra 69.2%
- ITBench Not in model cards Agents handle IT automation tasks in Kubernetes environments: site reliability incidents, compliance and security checks, and cloud cost work. Gemini 3 Pro 58.3%