Best AI Models for Coding.
Best models for code generation and debugging.
| # | Model | DFO Score ▾ | Tok/s | In $/1M | Out $/1M | Ctx | |
|---|---|---|---|---|---|---|---|
| 1 |  Claude Fable 5Anthropic | TOP PICK | 91.5 | 56.5 | $10.00 | $50.00 | 1M |
| 2 | Claude Opus 4.8Anthropic | TOP PICKIN-HOUSE PICK | 89.3 | 52.4 | $5.00 | $25.00 | 1M |
| 3 |  Gemini 3.1 Pro PreviewGoogle | FRONTIER | 85.4 | 126 | $2.00 | $12.00 | 1M |
| 4 |  GLM 5.2Z.ai | BEST VALUE | 84.7 | 167 | $0.2338 | $0.7348 | 1M |
| 5 | Claude Sonnet 5Anthropic | IN-HOUSE PICKNEW | 84.6 | 82.8 | $2.00 | $10.00 | 1M |
| 6 |  Grok 4.5xAI | BEST FOR CODINGNEW | 84.3 | 73.8 | $2.00 | $6.00 | 500K |
| 7 |  GPT-5.5OpenAI | BEST FOR CODING | 83.3 | 68.5 | $5.00 | $30.00 | 1.1M |
| 8 |  Qwen3.5 397B A17BQwen | 82.0 | 60.7 | $0.3900 | $2.34 | 262K | |
| 9 | Grok 4 FastxAI | 82.0 | 76.7 | $0.2000 | $0.5000 | 2M | |
| 10 | Gemini 2.5 Pro Preview 05-06Google | 82.0 | – | $1.25 | $10.00 | 1M | |
| 11 |  KAT-Coder-Pro V2Kwaipilot | 82.0 | 103 | $0.3000 | $1.20 | 256K | |
| 12 |  DeepSeek-Coder-V2DeepSeek | 82.0 | – | Free | Free | – | |
| 13 | DeepSeek V3.2 SpecialeDeepSeek | 82.0 | – | $0.2870 | $0.4310 | 164K | |
| 14 | gpt-oss-20bOpenAI | 82.0 | 220 | $0.0300 | $0.1300 | 131K | |
| 15 | QwQ 32BQwen | 82.0 | 31.9 | $0.1500 | $0.5800 | 131K | |
| 16 | GPT-5.5 (medium)OpenAI | 82.0 | 64.4 | $5.00 | $30.00 | – | |
| 17 | KAT-Coder-Pro V1Kwaipilot | 82.0 | 109 | $0.2070 | $0.8280 | 256K | |
| 18 |  Devstral Small 1.1Mistral | 82.0 | 47.5 | $0.1000 | $0.3000 | 131K | |
| 19 | Qwen2.5 72B InstructQwen | 82.0 | 55.7 | $0.3600 | $0.4000 | 131K | |
| 20 |  ERNIE 5.0 Thinking PreviewBaidu | 82.0 | – | Free | Free | – |
We deploy these models for businesses every week. Get a recommendation for your workload.
Get StartedBest models for code generation and debugging.
Leaderboards by use case
The overall table, re-ranked for the job you're hiring a model for.
How we rank AI models
The Design for Online AI Model Leaderboard scores 612 models on a single 0ā100 scale built from four weighted dimensions: intelligence (reasoning and knowledge benchmarks), technical capability (coding and tool use), content quality (writing and instruction-following) and value (capability per dollar).
Underlying data is aggregated from the OpenRouter API for pricing and availability, Artificial Analysis for intelligence, coding and agentic indices, and the Hugging Face Open LLM Leaderboard for open-model benchmarks. The fourth source is our own: we deploy these models in client agents, chatbots and automations every week, and that internal testing feeds the editorial layer, so a model that benchmarks well but is impractical to deploy will not automatically top the table.
Models are grouped into tiers (Frontier, Professional, Specialist, Efficient, Emerging and Legacy) to make like-for-like comparison easier, and newly released models are flagged so you can see what has just landed.
Leaderboard FAQ
How often is the leaderboard updated?
Pricing, availability and benchmark data are synced daily from our sources, and editorial scores are reviewed whenever a significant new model is released.
How is the overall score calculated?
Each model is graded 0ā10 on intelligence, technical capability, content quality and value; those dimensions are weighted and combined into the 0ā100 overall score used to rank the table.
Where does the data come from?
From four sources: the OpenRouter API, Artificial Analysis, the Hugging Face Open LLM Leaderboard, and internal testing from real deployments by the Design for Online team.
