Best AI model for agents
The best AI model for agents on current benchmarks is Claude Opus 5.5 (Anthropic): as of November 9, 2026 it has the highest Terminal-Bench 4.0 score TensorFeed tracks, 65.2%, tied with Claude Fable 5.2, among 45 models scored. The top four sit within 1.5 points of each other. It also leads FrontierCode v1.1 (54.6%).
Ranked on Terminal-Bench 4.0. Benchmark data as of November 9, 2026; prices as of November 9, 2026.
Terminal-Bench 4.0: 64.14%. $2.00 / $10.00 per 1M tokens.
Ranking on Terminal-Bench 4.0
Each score links to its source; vendor-reported and independent figures are labeled and are not interchangeable.
| # | Model | Provider | Terminal-Bench 4.0 | Price in / out | TFII | FrontierCode v1.1 | OSWorld 2.0 |
|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | 65.15%independent | $4.00 / $20.00 | 94.7 | 54.6% | - |
| 2 | Claude Fable 5.2 | Anthropic | 65.15%independent | $8.00 / $40.00 | - | - | - |
| 3 | Claude Sonnet 5.5 | Anthropic | 64.14%independent | $2.00 / $10.00 | 91.4 | 52.1% | - |
| 4 | Kimi K3.1 Pro | Moonshot AI | 63.64%independent | $2.00 / $10.00 | - | - | - |
| 5 | Gemini 4 Carbon | 63.13%independent | - | - | - | - | |
| 6 | GPT-6 Astra | OpenAI | 59.6%independent | $10.00 / $50.00 | 81.7 | 53.3% | - |
| 7 | Claude Fable 5.1 | Anthropic | 58.08%independent | $10.00 / $50.00 | 87.8 | 50.9% | - |
| 8 | Gemini 4 Argon | 57.58%independent | $2.00 / $10.00 | - | - | - | |
| 9 | GPT-6.1 Sol | OpenAI | 55.05%independent | $2.00 / $10.00 | - | 50.2% | - |
| 10 | Claude Opus 5 | Anthropic | 53.54%independent | $5.00 / $25.00 | 82.3 | 53.4% | 70.6% |
| 11 | Pareto | Unbiased | 50.8%vendor | $2.50 / $7.50 | - | - | - |
| 12 | Muse Wonder | Meta | 49.49%independent | $4.00 / $20.00 | - | - | - |
| 13 | GPT-6.1 Terra | OpenAI | 49.49%independent | $1.00 / $5.00 | - | - | - |
| 14 | GPT-6 Sol | OpenAI | 44.44%independent | $2.00 / $10.00 | - | 49.3% | - |
| 15 | Claude Fable 5 | Anthropic | 41.41%independent | $10.00 / $50.00 | 80.0 | 53.5% | 66.1% |
| 16 | GLM-5.3 | Z.ai | 38.89%independent | $1.40 / $4.40 | 69.6 | 40.1% | - |
| 17 | LongCat-2.5 | Meituan | 38.5%vendor | - | 66.3 | - | - |
| 18 | GPT-5.6 Sol | OpenAI | 37.88%independent | $4.00 / $20.00 | 74.4 | 47.5% | 62.6% |
| 19 | GLM-5.3 Flash | Z.ai | 37.37%independent | $0.15 / $0.50 | 60.8 | 31.8% | - |
| 20 | Claude Haiku 5.5 | Anthropic | 36.36%independent | $0.10 / $0.50 | 67.7 | 46.4% | - |
| 21 | Qwen3.8-Max | Alibaba | 34.34%independent | $2.00 / $6.00 | 77.4 2024 set | - | - |
| 22 | MiMo-V2.6-Pro | Xiaomi | 31.31%independent | $0.43 / $0.87 | - | - | - |
| 23 | Grok 4.7 | xAI | 28.79%independent | $2.00 / $6.00 | - | - | - |
| 24 | Mistral Large 4 | Mistral | 26.26%independent | $0.70 / $2.10 | - | - | - |
| 25 | MiniMax M3.1 Flash | MiniMax | 25.25%independent | $0.15 / $0.60 | - | - | - |
| 26 | MiMo-V2.6-Flash | Xiaomi | 24.24%independent | $0.14 / $0.28 | - | - | - |
| 27 | Claude Opus 4.8 | Anthropic | 23.23%independent | $5.00 / $25.00 | 66.5 | 46.5% | 55.7% |
| 28 | GPT-5.6 Terra | OpenAI | 22.73%independent | $2.00 / $12.00 | 60.8 | 41.3% | 50.2% |
| 29 | Ember-1 | Fireworks AI | 20.71%independent | $3.00 / $15.00 | - | - | - |
| 30 | DeepSeek V4.1 Flash | DeepSeek | 19.7%independent | $0.30 / $1.20 | - | - | - |
| 31 | Gemini 3.8 Flash | 19.7%independent | $0.75 / $3.75 | 79.0 2024 set | 41.2% | - | |
| 32 | Grok 4.6 | xAI | 17.17%independent | $2.00 / $6.00 | 83.2 2024 set | 48% | - |
| 33 | Kimi K3 | Moonshot AI | 17.17%independent | $3.00 / $15.00 | 63.7 | 44.2% | 58.3% |
| 34 | GPT-6 Luna | OpenAI | 13.64%independent | $0.10 / $0.50 | - | 42.4% | - |
| 35 | Gemini 3.7 Flash | 12.12%independent | - | 52.3 | 43.6% | 47.9% | |
| 36 | GPT-5.6 Luna | OpenAI | 11.62%independent | $0.20 / $1.20 | 53.2 | 39.8% | 45.6% |
| 37 | Claude Sonnet 5 | Anthropic | 10.61%independent | $2.00 / $10.00 | 57.5 | 42.7% | - |
| 38 | Muse Spark 1.3 | Meta | 10.61%independent | $1.25 / $4.25 | - | - | - |
| 39 | Grok 4.5 | xAI | 8.59%independent | $2.00 / $6.00 | 80.0 2024 set | 42.4% | - |
| 40 | Hy4 preview | Tencent | 8.08%independent | $0.83 / $2.50 | - | - | - |
| 41 | Muse Spark Open | Meta | 6.57%independent | $1.25 / $4.25 | - | - | - |
| 42 | Muse Spark 1.2 | Meta | 6.06%independent | $1.25 / $4.25 | - | - | - |
| 43 | Gemini 3.5 Flash | 6.06%independent | - | 77.8 2024 set | - | - | |
| 44 | MiniMax M3 | MiniMax | 1.01%independent | $0.30 / $1.20 | 29.7 | 14.7% | - |
| 45 | Inkling | Thinking Machines | 0.51%independent | - | 34.3 | 14% | - |
Prices are list prices per 1M input / output tokens. TFII is the headline TensorFeed Intelligence Index score; scores marked 2024 set come from older benchmarks and are not comparable with the rest.
How this is ranked
Terminal-Bench 4.0 is built from long, unattended jobs: 66 tasks across software, science, machine learning, operations, security and media, each producing a real deliverable and scored all or nothing. TensorFeed reports the Vals AI run, which uses one harness for every model and averages three runs, so the column compares like with like. That makes it the closest public proxy for whether an agent finishes work without a human stepping in.
The harness moves this number more than most model upgrades do. Scores here all come from the same Vals AI harness; figures from a lab's own agent, or from before the September 29, 2026 harness change, are not comparable.
The full board, scoring notes and score ranges are on the Terminal-Bench 4.0 leaderboard. The same scores are free as JSON at /api/benchmarks.
Questions
What is the best AI model for AI agents?
The best AI model for agents on current benchmarks is Claude Opus 5.5 (Anthropic): as of November 9, 2026 it has the highest Terminal-Bench 4.0 score TensorFeed tracks, 65.2%, tied with Claude Fable 5.2, among 45 models scored. The top four sit within 1.5 points of each other. It also leads FrontierCode v1.1 (54.6%).
Which of the top AI models for agents is cheapest?
Of the five highest Terminal-Bench 4.0 scores TensorFeed tracks, the lowest-priced model with a list price is Claude Sonnet 5.5 (64.1%), at $2.00 input and $10.00 output per 1M tokens as of November 9, 2026.
Why rank AI models for agents on Terminal-Bench 4.0?
Terminal-Bench 4.0 is built from long, unattended jobs: 66 tasks across software, science, machine learning, operations, security and media, each producing a real deliverable and scored all or nothing. TensorFeed reports the Vals AI run, which uses one harness for every model and averages three runs, so the column compares like with like. That makes it the closest public proxy for whether an agent finishes work without a human stepping in.