Best LLM for coding — ranked by your priorities
The best LLM for coding is the one that patches your repo at a latency and price you can live with — not the one that topped last month’s blog roundup. Coding quality moves with tool-calling, context (so the model sees the right files), and how often you retry.
On the leaderboard, raise quality and speed weights to surface strong coding models, or raise cost if you generate a lot of boilerplate. Compare finalists side-by-side, then stress-test them in the arena with prompts from your actual codebase.
Highest quality index right now
Live OpenRouter pricing and Arena quality, cached about an hour. Not a static blog table.
| Model | Provider | Blend $ / 1M | Quality | Context |
|---|---|---|---|---|
| Qwen3.8 Max (0902) | Qwen | 4.00 | 93.0 | 1,000,000 |
| Claude Opus 5 | Anthropic | 15.00 | 91.5 | 1,000,000 |
| Claude Fable 5 | Anthropic | 30.00 | 89.4 | 1,000,000 |
| GLM 5.3 Flash | Z.ai | 0.16 | 87.9 | 1,048,576 |
| Qwen3.8 27B | Qwen | 1.71 | 87.4 | 1,000,000 |
What “best at coding” usually means
People mix four jobs: inline completion, multi-file refactor, debugging from a stack trace, and agentic tool use (edit, test, repeat). A model that is great at single-file generation can be mediocre at agents. Decide which job is 80% of your tokens. Completion wants speed; agents want reliability and enough context to see tests.
Benchmarks vs your repo
HumanEval-style scores and Arena Elo are weak proxies for a private monolith. They still help you drop obviously weak rows. After a shortlist of three, paste a real failing test and a real module into the arena. Vote on whether the patch would merge, not on whether the prose sounds confident.
Context is a coding feature
If you stuff whole repos into the prompt, max context and input price dominate. If you use a retrieval or agent that opens a few files, quality and output price dominate. Compare context on the model pages; do not buy a 1M window for a 4k-token loop.
Cost of retries
A model that is 2× more expensive but succeeds in one try beats a cheap model you run four times. Log retries in your coding agent. If DeepSeek-class rows succeed 90% of the time on boilerplate and GPT/Claude win on architecture changes, route accordingly.
A one-hour selection process
Set leaderboard weights to quality 0.5, speed 0.25, cost 0.25. Take the top two closed and top one open-weight model. Compare them. Run two arena prompts: “fix this failing test” and “explain this module.” Ship the cheaper one that you would actually merge.
FAQ
What is the best coding LLM right now?
- It changes. Rank by quality and speed on the live leaderboard, then confirm with your own repo prompts in the arena.
Is Claude better than GPT for code?
- Sometimes, for review and long diffs; GPT often wins on tool ecosystems. Compare current SKUs, not brand memory.
Can I use open-weight models for coding?
- Yes. Filter open weights and compare quality vs cost. Self-hosting is a separate ops decision.
Should I use a reasoning model for coding?
- For hard bugs and architecture, try one. For autocomplete and nits, a fast cheap model is usually enough.
Related guides
Also try: LLM leaderboard, compare tool, live arena.