Terminal-Bench 2.0 (Vals)
Terminal-Bench 2.0 in Vals AI's fixed harness. Run by Vals AI.
As of 19 Sept 2026, GPT-5.5 leads Terminal-Bench 2.0 (Vals) on BenchLeader with 73.2%, ahead of Claude Opus 4.8 at 70.0%, across 64 model configurations with a published result.
- Published by
- Vals AI
- Category
- Agents & tools
- Index weight
- Reference only
- Models
- 64
- Data as of
- 19 Sept 2026
Vals AI (vals.ai).
What the test looks like
Terminal-Bench 2.0: agent tasks in a shell, checked by tests, run by Vals AI.
How it is scored
Percent of tasks passing, run by Vals AI.
What to keep in mind
Superseded by 2.1, kept for comparison.
- 1GPT-5.5 (high)73.2%
- 2Claude Opus 4.870.0%
- 3Claude Opus 4.768.5%
- 4Gemini 3.5 Flash (high)67.4%
- 5Gemini 3.1 Pro (high)67.4%
- 6GPT-5.3 Codex (high)64.0%
- 7Muse Spark59.5%
- 8Claude Sonnet 4.659.5%
- 9Qwen3 7 (max)59.2%
- 10Claude Opus 4.6 (thinking)58.4%
- 11GPT-5.4 (high)58.4%
- 12Claude Opus 4.558.4%
- 13Kimi K2.657.3%
- 14DeepSeek V4 Pro56.2%
- 15Gemini 3 Pro (high)55.1%
64 of 64
| # | ||||
|---|---|---|---|---|
| 1 |  GPT-5.5highOpenAI | 73.2% | 67.0 | 2026-06-04 |
| 2 |  Claude Opus 4.8Anthropic | 70.0% | 61.9 | 2026-06-04 |
| 3 | Claude Opus 4.7Anthropic | 68.5% | 64.5 | 2026-06-04 |
| 4 |  Gemini 3.5 FlashhighGoogle | 67.4% | 63.6 | 2026-06-04 |
| 5 | Gemini 3.1 ProhighGoogle | 67.4% | 60.7 | 2026-06-04 |
| 6 | GPT-5.3 CodexhighOpenAI | 64.0% | – | 2026-06-04 |
| 7 |  Muse SparkMeta | 59.5% | 65.7 | 2026-06-04 |
| 8 | Claude Sonnet 4.6Anthropic | 59.5% | 59.0 | 2026-06-04 |
| 9 |  Qwen3 7maxAlibaba | 59.2% | 61.0 | 2026-06-04 |
| 10 | Claude Opus 4.6thinkingAnthropic | 58.4% | 62.5 | 2026-06-04 |
| 11 | GPT-5.4highOpenAI | 58.4% | 59.0 | 2026-06-04 |
| 12 | Claude Opus 4.5Anthropic | 58.4% | 58.2 | 2026-06-04 |
| 13 |  Kimi K2.6Moonshot AIopen ↗ | 57.3% | 60.5 | 2026-06-04 |
| 14 |  DeepSeek V4 ProDeepSeekopen ↗ | 56.2% | 52.6 | 2026-06-04 |
| 15 | Gemini 3 ProhighGoogle | 55.1% | 61.1 | 2026-06-04 |
| 16 | Claude Opus 4.5thinkingAnthropic | 53.9% | 60.8 | 2026-06-04 |
| 17 |  GLM-5.1Zhipu AIopen ↗ | 53.9% | 56.9 | 2026-06-04 |
| 18 | Qwen3 6maxAlibaba | 51.7% | 63.3 | 2026-06-04 |
| 19 | Gemini 3 FlashhighGoogle | 51.7% | 58.7 | 2026-06-04 |
| 20 | GPT-5.2highOpenAI | 51.7% | 55.7 | 2026-06-04 |
| 21 | GLM-5thinkingZhipu AIopen | 49.4% | 59.6 | 2026-06-04 |
| 22 |  MiniMax-M2.7MiniMaxopen ↗ | 47.2% | 56.1 | 2026-06-04 |
| 23 | MiniMax-M3MiniMaxopen ↗ | 46.1% | 57.5 | 2026-06-04 |
| 24 | GPT-5.1highOpenAI | 44.9% | 58.7 | 2026-06-04 |
| 25 | Qwen3.6 PlusAlibabaopen | 44.9% | 58.5 | 2026-06-04 |
| 26 | GPT-5.4 minihighOpenAI | 44.9% | 53.0 | 2026-06-04 |
| 27 | Qwen3.6 27BAlibaba | 44.9% | 44.7 | 2026-06-04 |
| 28 |  Grok 4.3xAI | 43.5% | 50.5 | 2026-06-04 |
| 29 | Qwen3.5 PlusthinkingAlibaba | 41.6% | 57.3 | 2026-06-04 |
| 30 | MiniMax-M2.5MiniMaxopen ↗ | 41.6% | 54.2 | 2026-06-04 |
| 31 | Claude Sonnet 4.5thinkingAnthropic | 41.6% | 53.4 | 2026-06-04 |
| 32 | Grok 4.20thinkingxAI | 40.5% | 59.9 | 2026-06-04 |
| 33 | Kimi K2.5thinkingMoonshot AIopen | 40.5% | 58.5 | 2026-06-04 |
| 34 | GPT-5.4 nanohighOpenAI | 39.9% | 49.1 | 2026-06-04 |
| 35 | Gemma 4 31BhighGoogleopen ↗ | 39.3% | – | 2026-06-04 |
| 36 | GLM-4.7Zhipu AIopen ↗ | 38.2% | 52.4 | 2026-06-04 |
| 37 | Claude Haiku 4.5thinkingAnthropic | 38.2% | 47.5 | 2026-06-04 |
| 38 | GPT-5highOpenAI | 37.1% | 58.6 | 2026-06-04 |
| 39 | MiniMax-M2MiniMaxopen | 37.1% | 55.1 | 2026-06-04 |
| 40 | Kimi K2 ThinkingthinkingMoonshot AIopen | 37.1% | 52.2 | 2026-06-04 |
| 41 | DeepSeek V3.2highDeepSeekopen | 36.0% | – | 2026-06-04 |
| 42 |  Laguna M.1Poolsideopen | 31.5% | 37.7 | 2026-06-04 |
| 43 | Gemini 2.5 ProGoogle | 30.3% | 54.5 | 2026-06-04 |
| 44 |  Mistral Medium 3.5highMistral AIopen | 30.3% | – | 2026-06-04 |
| 45 | Grok 4 FastthinkingxAI | 29.2% | 52.6 | 2026-06-04 |
| 46 | Grok 4xAI | 28.1% | 58.5 | 2026-06-04 |
| 47 | GLM-4.6Zhipu AIopen | 28.1% | 49.8 | 2026-06-04 |
| 48 | Laguna XS.2Poolside | 28.1% | 36.5 | 2026-06-04 |
| 49 | GPT-5 minihighOpenAI | 27.0% | 55.3 | 2026-06-04 |
| 50 | Kimi K2Moonshot AIopen | 25.8% | 51.0 | 2026-06-04 |
| 51 | Grok 4.1thinkingxAI | 24.7% | 56.0 | 2026-06-04 |
| 52 | Qwen3 MaxmaxAlibaba | 24.7% | 53.5 | 2026-06-04 |
| 53 | Qwen3.5 FlashAlibaba | 24.7% | 52.5 | 2026-06-04 |
| 54 | Gemini 3.1 Flash LitehighGoogle | 24.7% | 49.1 | 2026-06-04 |
| 55 | Gemini 2.5 Flash 09 2025thinkingGoogle | 21.4% | 51.6 | 2026-06-04 |
| 56 | gpt-oss-120bOpenAIopen | 19.1% | 46.9 | 2026-06-04 |
| 57 |  Trinity LargethinkingArcee AIopen | 18.0% | 47.6 | 2026-06-04 |
| 58 | Grok 4.1no reasoningxAI | 18.0% | 43.0 | 2026-06-04 |
| 59 | Mistral Small 4highMistral AIopen ↗ | 16.9% | – | 2026-06-04 |
| 60 | GPT-4.1highOpenAI | 14.6% | 46.6 | 2026-06-04 |
Cite as: BenchLeader, “Terminal-Bench 2.0 (Vals) leaderboard”, https://www.benchleader.com/benchmarks/vals_terminal_bench_2, data as of 19 Sept 2026.
Terminal-Bench 2.0 (Vals): questions
- What does Terminal-Bench 2.0 (Vals) measure?
- Terminal-Bench 2.0: agent tasks in a shell, checked by tests, run by Vals AI. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads Terminal-Bench 2.0 (Vals)?
- GPT-5.5 leads Terminal-Bench 2.0 (Vals) with 73.2% as of 19 Sept 2026, ahead of Claude Opus 4.8 at 70.0%.
- How many models have Terminal-Bench 2.0 (Vals) results?
- 64 model configurations have a Terminal-Bench 2.0 (Vals) result on BenchLeader, all taken from Vals AI.
- Who runs Terminal-Bench 2.0 (Vals) and how often is it updated?
- Terminal-Bench 2.0 (Vals) is published by Vals AI. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does Terminal-Bench 2.0 (Vals) count toward the BenchLeader Index?
- No. Terminal-Bench 2.0 (Vals) is shown for reference but left out of the composite index.