Best AI Models for Coding Agents
Ranking of language models for autonomous coding agents — the backbone of tools like Cursor, Claude Code, Windsurf and Devin.Updated August 2026. Based on independent benchmark data.
Why This Matters
Choosing the right model for coding agents is about more than raw benchmark scores. The models below are ranked using a weighted blend of category-specific benchmarks, arena Elo, and the intelligence index — but real-world performance depends on your context window, latency budget, and integration path.
For coding agents, we weighted the benchmarks that most strongly predict success in production: task accuracy first, then speed and cost efficiency. Leaders on this page consistently outperform on both standardized tests and real-world workloads.
If you operate under a strict budget, check the Best Value LLMs ranking. If latency matters more than raw quality, see Fastest LLMs. Both pages use the same underlying dataset with different weightings.
| # | Model | Provider | Score | Speed | Price (output) |
|---|---|---|---|---|---|
| 1 | Claude Fable 5 | Anthropic | 91.4 | 71 t/s | $50/M |
| 2 | Kimi K3 | Moonshot AI | 84.6 | 37 t/s | $15/M |
| 3 | GPT-6 Astra | OpenAI | 70.4 | 82 t/s | $50/M |
| 4 | Claude Opus 4.8 | Anthropic | 68.1 | 50 t/s | $15/M |
| 5 | Gemini 3.5 Flash | 65.4 | 267 t/s | $9/M | |
| 6 | Claude Opus 4.7 | Anthropic | 64.3 | 80 t/s | $15/M |
| 7 | Qwen3.8 Flash Next | Alibaba | 62.3 | 168 t/s | $0.4/M |
| 8 | GLM-5.3 Flash | Z.ai | 60.0 | 182 t/s | $0.3/M |
| 9 | Qwen3.7 Max | Alibaba | 59.3 | 203 t/s | $7.5/M |
| 10 | MiMo V2.5 Pro | Xiaomi | 58.4 | 65 t/s | $0.87/M |
| 11 | GPT-5.5 | OpenAI | 58.3 | 60 t/s | $10/M |
| 12 | Granite 4.2 30B Instruct | IBM | 57.2 | 78 t/s | $1.2/M |
| 13 | GLM-5.2 | Z.ai | 56.9 | 164 t/s | $4.29/M |
| 14 | Gemini 3.1 Pro | 54.9 | 136 t/s | $12/M | |
| 15 | DeepSeek V4.1 Flash | DeepSeek | 49.6 | 214 t/s | $1.2/M |
| 16 | Qwen3.7 Plus | Alibaba | 47.5 | 54 t/s | $1.6/M |
| 17 | GPT-5.6 Sol | OpenAI | 46.3 | 72 t/s | $30/M |
| 18 | GPT-5.6 Luna | OpenAI | 44.2 | 175 t/s | $1.2/M |
| 19 | Granite 4.2 8B Instruct | IBM | 42.8 | 115 t/s | $0.45/M |
| 20 | DeepSeek V4 Pro | DeepSeek | 42.1 | 66 t/s | $0.87/M |
Top 5 Coding Agents — Quick Comparison
Methodology
Rankings are based on independent benchmark evaluations. The coding agents score is computed from multiple standardized benchmarks and normalized to 0–100. Data is updated regularly from public benchmark results.
Related Rankings
Frequently Asked Questions
What is the best AI model for coding agents like Cursor or Claude Code?
Claude and GPT-class models lead agentic coding benchmarks like SWE-bench Pro and TAU-bench, which measure end-to-end software task completion.