Best for: Coding
Best LLM for Coding
Ranked on SWE-Bench, HumanEval, and dollars-per-1M output tokens. Balanced for autonomous and assistive coding workflows.
Updated May 2026. Top 3 this month: GPT-4o (2024-11-20), Claude Sonnet 4.5, GPT-5 Codex.
Podium
This month’s top three.
- 1GPT-4o (2024-11-20)OpenAI
- Input / 1M
- $2.50
- Output / 1M
- $10.00
- Context
- 128,000
- 2Claude Sonnet 4.5Anthropic
- Input / 1M
- $3.00
- Output / 1M
- $15.00
- Context
- 1,000,000
- 3GPT-5 CodexOpenAI
- Input / 1M
- $1.25
- Output / 1M
- $10.00
- Context
- 400,000
How we rank
Weights tuned for coding.
Choosing an LLM for coding comes down to three things: how well it turns specifications into working code, how well it reasons about large repositories, and how much it will cost once you wire it into CI or an agent loop. We weight SWE-Bench heaviest because it best predicts real-world coding-agent success, followed by HumanEval for short-form correctness, and a price pillar so the recommendation survives contact with a finance review.
Our full methodology is published on the methodology page.
Pillars and weights:
- SWE-Bench50%
- HumanEval30%
- price20%
Full ranking
Top ranked models
| Rank | Model | Provider | Input $/1M | Output $/1M | Context |
|---|---|---|---|---|---|
| 1 | GPT-4o (2024-11-20) | OpenAI | $2.50 | $10.00 | 128,000 |
| 2 | Claude Sonnet 4.5 | Anthropic | $3.00 | $15.00 | 1,000,000 |
| 3 | GPT-5 Codex | OpenAI | $1.25 | $10.00 | 400,000 |
| 4 | Gemini 2.5 Pro | $1.25 | $10.00 | 1,048,576 | |
| 5 | Gemini 2.5 Pro Preview 06-05 | $1.25 | $10.00 | 1,048,576 | |
| 6 | GPT-5.1-Codex | OpenAI | $1.25 | $10.00 | 400,000 |
| 7 | o3 | OpenAI | $2.00 | $8.00 | 200,000 |
| 8 | Claude 3.7 Sonnet | Anthropic | $3.00 | $15.00 | 200,000 |
| 9 | Claude 3.7 Sonnet (thinking) | Anthropic | $3.00 | $15.00 | 200,000 |
| 10 | GPT-5 Mini | OpenAI | $0.25 | $2.00 | 400,000 |
Field notes
Tips for coding
- 01
Prefer a model with a large context window if your repo is bigger than ~200 files.
- 02
Use batch pricing for CI / nightly refactor jobs; interactive IDE work stays on the standard price.
- 03
Check function-calling reliability before committing to an agentic flow.
FAQ
Frequently asked questions
The questions teams ask before picking a model for coding.
Get instant answers from our AI agent
As of May 2026, our weighted top 3 are GPT-4o (2024-11-20), Claude Sonnet 4.5, GPT-5 Codex.
Claude wins long-horizon refactoring; GPT wins short-burst correctness. The right answer depends on your workload mix — see the scoring pillars below.
Rarely. Fine-tuning on proprietary code still helps, but for 90% of shops a strong frontier model with RAG over the repo gets you most of the way.
DeepSeek and Meta Llama variants are competitive on price. We list their hosted pricing here; self-host economics live in our Shadow AI audit tool.