| # | Model | Provider | Score | Evidence |
|---|---|---|---|---|
| 1 |  MiniMax M2.5 | MiniMax | 75.8 | 1 row · 20% of weighting |
| 2 |  Gemini 3 Flash Preview | 75.8 | 1 row · 20% of weighting | |
| 3 |  Claude Opus 4.6 | Anthropic | 75.6 | 1 row · 20% of weighting |
| 4 | Claude Opus 4.5 | Anthropic | 74.4 | 1 row · 20% of weighting |
| 5 |  GLM 5 | Z.ai | 72.8 | 1 row · 20% of weighting |
| 6 |  GPT-5.2-Codex | OpenAI | 72.8 | 1 row · 20% of weighting |
| 7 | GPT-5.2 | OpenAI | 72.8 | 1 row · 20% of weighting |
| 8 | GPT-5 | OpenAI | 71.8 | 1 row · 20% of weighting |
| 9 |  Kimi K2.5 | Moonshot AI | 70.8 | 1 row · 20% of weighting |
| 10 | Claude Sonnet 4 | Anthropic | 70.8 | 1 row · 20% of weighting |
| 11 | Claude Sonnet 4.5 | Anthropic | 70.6 | 1 row · 20% of weighting |
| 12 |  DeepSeek V3.2 | DeepSeek | 70.0 | 1 row · 20% of weighting |
| 13 | Claude Haiku 4.5 | Anthropic | 66.6 | 1 row · 20% of weighting |
| 14 | GPT-5.1-Codex | OpenAI | 66.0 | 1 row · 20% of weighting |
| 15 | GPT-5.1 | OpenAI | 66.0 | 1 row · 20% of weighting |
| 16 | o1 | OpenAI | 64.6 | 1 row · 20% of weighting |
| 17 | Kimi K2 Thinking | Moonshot AI | 63.4 | 1 row · 20% of weighting |
| 18 | MiniMax M2 | MiniMax | 61.0 | 1 row · 20% of weighting |
| 19 |  Qwen3 Coder 30B A3B Instruct | Qwen | 60.4 | 1 row · 20% of weighting |
| 20 | GPT-5 Mini | OpenAI | 59.8 | 1 row · 20% of weighting |
| 21 | o3 | OpenAI | 58.4 | 1 row · 20% of weighting |
| 22 | GLM 4.6 | Z.ai | 55.4 | 1 row · 20% of weighting |
| 23 | GLM 4.5 | Z.ai | 54.2 | 1 row · 20% of weighting |
| 24 | Gemini 2.5 Pro | 53.6 | 1 row · 20% of weighting | |
| 25 | o4 Mini | OpenAI | 45.0 | 1 row · 20% of weighting |
| 26 | o3 Mini | OpenAI | 42.4 | 1 row · 20% of weighting |
| 27 | DeepSeek V3 0324 | DeepSeek | 42.0 | 1 row · 20% of weighting |
| 28 | GPT-4.1 | OpenAI | 39.6 | 1 row · 20% of weighting |
| 29 | Qwen2.5 Coder 32B Instruct | Qwen | 38.0 | 1 row · 20% of weighting |
| 30 | GPT-5 Nano | OpenAI | 34.8 | 1 row · 20% of weighting |
| 31 | Gemini 2.5 Flash | 28.7 | 1 row · 20% of weighting | |
| 32 | GPT-4o (2024-11-20) | OpenAI | 26.2 | 1 row · 20% of weighting |
| 33 | gpt-oss-120b | OpenAI | 26.0 | 1 row · 20% of weighting |
| 34 | GPT-4.1 Mini | OpenAI | 23.9 | 1 row · 20% of weighting |
| 35 | GPT-4 | OpenAI | 22.4 | 1 row · 20% of weighting |
How the leaderboard works
- How is the LLM leaderboard ranked?
- Each model is scored per category from its published benchmark results, then the categories are weighted into one figure: agentic 22%, coding 20%, reasoning 17%, multimodal 12%, knowledge 12%, multilingual 7%, instruction following 5%, maths 5%. A category with no published evidence is left out and the rest re-weighted, never counted as zero.
- Why isn't every model on the leaderboard?
- A model with no published benchmark score is untested, not last, so it does not appear. Every tracked model still has its own page, with its price, context window and every update, whether or not it has been scored.
- What does the coverage column mean?
- It is how much of the category weighting a model's score rests on. A score built on 90% of the weighting is a stronger claim than one built on 20%, so both are shown side by side rather than hiding the difference.
- How often does the leaderboard change?
- Scores are added as benchmark results are published and read, and model catalogues are re-read every three hours, so a new model or a new result can move the ranking the same day.
Are these models citing you, or your competitors?
WriteWorks tracks how often 10+ AI platforms, from ChatGPT and Claude to Gemini and Perplexity, mention your brand against your competitors, then helps you engineer the content that changes the answer.