BlogLLM Benchmark Scores as Published, Side by Side
Arena-Hard: Results as Published
Every result the catalog holds for Arena-Hard: 16 models from 7 developers, read from 4 pages. Each number is as printed by the company named beside it; nothing is ranked or rescored. Models are listed by developer, newest first.
Area Preference and Composite RankingsMaintainer LMArena (Tianle Li and co-authors)Introduced 2024Data version 2026103001, as of 2026-10-30
Arena-Hard tests open-ended answering on 500 hard prompts drawn from crowdsourced conversations. An LLM judge compares the model's answers with a baseline model's answers, and the score is the resulting win rate.
Arena-Hard win rate (GPT-4-1106 judge)
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| o1-miniOpenAI, released 2024-09-12 | 92% | DeepSeek, DeepSeek-R1 model cardchecked 2026-10-15 | - |
| GPT-4oOpenAI, released 2024-05-13 | 80.4% | DeepSeek, DeepSeek-R1 model cardchecked 2026-10-15; printed on 1 more page | - |
| Claude 3.7 SonnetAnthropic, released 2025-02-24 | 93.2% | Mistral AI, Medium is the new largechecked 2026-10-15 | - |
| Claude 3.5 SonnetAnthropic, released 2024-10-22 | 85.2% | DeepSeek, DeepSeek-R1 model cardchecked 2026-10-15 | - |
| Llama 4 MaverickMeta, released 2025-04-05 | 91.8% | Mistral AI, Medium is the new largechecked 2026-10-15 | - |
| DeepSeek-V3-0324DeepSeek, released 2025-03-24 | 97.3% | Mistral AI, Medium is the new largechecked 2026-10-15 | - |
| DeepSeek-R1DeepSeek, released 2025-01-20 | 92.3% | DeepSeek, DeepSeek-R1 model cardchecked 2026-10-15 | - |
| DeepSeek-V3DeepSeek, released 2024-12-26 | 85.5% | DeepSeek, DeepSeek-R1 model cardchecked 2026-10-15 | - |
| Mistral Small 4Mistral AI, released 2026-03-16 | 55.8% | Mistral AI, Mistral Small 4 model cardchecked 2026-10-15 | Effort: none |
| 58.3% | Mistral AI, Mistral Small 4 model cardchecked 2026-10-15 | Effort: high | |
| Mistral Large 3Mistral AI, released 2025-12-02 | 66.7% | Mistral AI, Mistral Small 4 model cardchecked 2026-10-15 | - |
| Mistral Medium 3.1Mistral AI, released 2025-08-12 | 67.3% | Mistral AI, Mistral Small 4 model cardchecked 2026-10-15 | - |
| Mistral Small 3.2Mistral AI, released 2025-06-20 | 43.1% | Mistral AI, Mistral Small 4 model cardchecked 2026-10-15 | - |
| Mistral Medium 3Mistral AI, released 2025-05-07 | 97.1% | Mistral AI, Medium is the new largechecked 2026-10-15 | - |
Arena-Hard v2 (average)
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| gpt-oss-20bOpenAI, released 2025-08-05 | 48.6% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | - |
| Qwen3-30B-A3B-Thinking-2507Alibaba, released 2025-07-30 | 57.8% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | Effort: thinking |
| Nemotron 3 Nano 30B-A3BNVIDIA, released 2025-12-15 | 67.7% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | - |
Arena-Hard v2 (creative writing)
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| gpt-oss-20bOpenAI, released 2025-08-05 | 25.9% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | - |
| Qwen3-30B-A3B-Thinking-2507Alibaba, released 2025-07-30 | 66% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | Effort: thinking |
| Nemotron 3 Nano 30B-A3BNVIDIA, released 2025-12-15 | 63.2% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | - |
Arena-Hard v2 (hard prompt)
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| gpt-oss-20bOpenAI, released 2025-08-05 | 71.2% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | - |
| Qwen3-30B-A3B-Thinking-2507Alibaba, released 2025-07-30 | 49.6% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | Effort: thinking |
| Nemotron 3 Nano 30B-A3BNVIDIA, released 2025-12-15 | 72.1% | NVIDIA, NVIDIA Nemotron 3 Nano 30B-A3B model cardchecked 2026-10-15 | - |
Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.