Best LLM for terminal and DevOps tasks
- Claude Opus 5.5 leads
- 65.2%
- models ranked
- 48
- lead over #2
- +1.0 pts
Terminal-Bench 4 (Vals AI) · read October 21
# Terminal-Bench 4 (Vals AI), read October 21
leaders = {
"Claude Opus 5.5": 65.2%, "Claude Sonnet 5.5": 64.1%, "GPT-6 Astra": 59.6%,
}
As of October 21, 2026, Anthropic Claude Opus 5.5 leads 48 models on Terminal-Bench 4 (Vals AI) with 65.2%, ahead of Anthropic Claude Sonnet 5.5 (64.1%) and OpenAI GPT-6 Astra (59.6%).
This list narrows the best LLM for coding ranking.
| Rank | Model | Terminal-Bench 4 (Vals AI) | Coding evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 Anthropic | 65.2% | 92.2 | $20.00 | |
| 2 | Claude Sonnet 5.5 Anthropic | 64.1% | 88.9 | $10.00 | |
| 3 | GPT-6 Astra OpenAI | 59.6% | 88.7 | $50.00 | |
| 4 | Claude Fable 5.1 Anthropic | 58.1% | 89.1 | $50.00 | |
| 5 | Gemini 4 Argon Google | 57.6% | 80.5 | $10.00 | |
| 6 | GPT-6.1 Sol OpenAI | 55.0% | 88.7 | $10.00 | |
| 7 | Claude Opus 5 Anthropic | 53.5% | 81.8 | $25.00 | |
| 8 | Muse Wonder Meta | 49.5% | 64.5 | $20.00 | |
| 9 | GPT-6 Sol OpenAI | 44.4% | 82.8 | $10.00 | |
| 10 | Claude Fable 5 Anthropic | 41.4% | 82.1 | $50.00 | |
| 11 | GLM 5.3 Z.ai | 38.9% | 73.1 | $1.40 | |
| 12 | GPT-5.6 Sol OpenAI | 37.9% | 76.7 | $20.00 | |
| 13 | GLM 5.3 Flash Z.ai | 37.4% | 71.0 | $0.50 | |
| 14 | Claude Haiku 5.5 Anthropic | 36.4% | 75.9 | $0.50 | |
| 15 | Qwen3.8 Max (0902) Qwen | 34.3% | 63.5 | $6.00 | |
| 16 | DeepSeek V4.1 Flash 1017 DeepSeek | 31.8% | 83.3 | $0.60 | |
| 17 | Step 5 Preview StepFun | 31.8% | 71.4 | $2.70 | |
| 18 | MiMo-V2.6-Pro Xiaomi | 31.3% | 72.6 | $0.87 | |
| 19 | Grok 4.7 xAI | 28.8% | 75.0 | $6.00 | |
| 20 | Mistral Large 4 Mistral AI | 27.8% | 55.3 | $4.18 |
How this list is built
Generated from the dataset on every refresh. Nobody picks the order by hand.
Models by their Terminal-Bench 4 (Vals AI) result, as the board publishes it, best first. Tasks finished in a real terminal with one shared harness, run by Vals AI. Share completed.
- OpenRouter, fetched
- Vals AI boards, fetched