Ranked by HumanEval, the canonical benchmark for code generation. Measures how often the model writes a Python function that passes hidden unit tests on the first try.
Data updated: August 11, 2026
| # | Model | Vendor | Arena Elo | SWE-bench | Price in/out ($/M) | Context |
|---|---|---|---|---|---|---|
| 1 | Claude Mythos | Anthropic | 1478 | 93.9% | $25 / $125 | 1M |
| 2 | Claude Fable 5 | Anthropic | 1492 | 77% | $10 / $50 | 1M |
| 3 | Claude Mythos 5 | Anthropic | 1493 | 78% | $10 / $50 | 1M |
| 4 | Kimi K3 | Moonshot AI | 1478 | 75.5% | $3 / $15 | 1M |
| 5 | GPT-5.4 Codex | OpenAI | 1408 | 70% | $9 / $36 | 500K |
| 6 | Codex CLI | OpenAI | 1408 | 70% | $9 / $36 | 500K |
| 7 | GPT-5 Codex | OpenAI | 1395 | 68% | $8 / $32 | 400K |
| 8 | Claude Opus 4.8 | Anthropic | 1435 | 67% | $5 / $25 | 1M |
| 9 | GPT-5.5 | OpenAI | 1432 | 66% | $5 / $30 | 600K |
| 10 | OpenAI o4 | OpenAI | 1438 | 72% | $12 / $48 | 256K |
| 11 | Claude Opus 4.7 | Anthropic | 1420 | 64.3% | $5 / $25 | 1M |
| 12 | Claude Code | Anthropic | 1420 | 64.3% | $15 / $75 | 1M |
| 13 | Gemini 3 Deep Think | Google DeepMind | 1429 | 64% | $14 / $56 | 1M |
| 14 | OpenAI o3 | OpenAI | 1418 | 69.1% | $2 / $8 | 200K |
| 15 | GPT-5 | OpenAI | 1412 | 65% | $1.25 / $10 | 400K |
The ranking above sorts every model by HumanEval. Frontier Claude and GPT models lead, and agentic benchmarks like SWE-bench separate models that fix real bugs in real repositories from models that only autocomplete. For day-to-day work, pick the top model your budget allows and route easy tasks to its smaller sibling.
Open-weight models like DeepSeek and Llama can be self-hosted for the cost of hardware, and most vendors offer free tiers of their chat products. For serious coding the paid APIs are usually worth it. The price of one hour of engineering time buys millions of tokens.
Both, for different things. HumanEval measures writing a correct function from scratch; SWE-bench Verified measures resolving a real GitHub issue in a large codebase. If you want an AI pair programmer, look at HumanEval; if you want an autonomous agent, SWE-bench is the number that matters.