Best LLM for research tasks
- Claude Opus 5 leads
- 88.9%
- models ranked
- 43
- lead over #2
- +2.0 pts
LiveBench research · read October 21
As of October 21, 2026, Anthropic Claude Opus 5 leads 43 models on LiveBench research with 88.9%, ahead of Anthropic Claude Fable 5 (86.9%) and DeepSeek V4.1 Flash (86.8%).
The full best LLM for agents ranking combines this benchmark with others.
| Rank | Model | LiveBench research | Coding evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | Claude Opus 5 Anthropic | 88.9% | 83.0 | $25.00 | |
| 2 | Claude Fable 5 Anthropic | 86.9% | 79.3 | $50.00 | |
| 3 | DeepSeek V4.1 Flash DeepSeek | 86.8% | 62.5 | $0.78 | |
| 4 | Grok 4.7 xAI | 86.4% | 66.7 | $6.00 | |
| 5 | Grok 4.5 xAI | 85.9% | 58.6 | $6.00 | |
| 6 | GPT-6 Astra OpenAI | 85.2% | 85.1 | $50.00 | |
| 7 | Claude Opus 5.5 Anthropic | 85.0% | 88.8 | $20.00 | |
| 8 | Grok 4.6 xAI | 84.5% | 65.7 | $6.00 | |
| 9 | Claude Sonnet 5.5 Anthropic | 83.6% | 87.3 | $10.00 | |
| 10 | GPT-6.1 Sol OpenAI | 83.5% | 82.4 | $10.00 | |
| 11 | Gemini 4 Argon Google | 83.1% | 81.0 | $10.00 | |
| 12 | Claude Sonnet 5 Anthropic | 82.8% | 57.7 | $10.00 | |
| 13 | Claude Fable 5.1 Anthropic | 82.0% | 84.2 | $50.00 | |
| 14 | Claude Haiku 5.5 Anthropic | 81.9% | 70.7 | $0.50 | |
| 15 | Claude Opus 4.8 Anthropic | 81.8% | 65.0 | $25.00 | |
| 16 | Muse Spark Open Meta | 80.3% | 47.4 | $4.25 | |
| 17 | DeepSeek V4.1 Flash 1017 DeepSeek | 79.5% | 70.3 | $0.60 | |
| 18 | Gemini 3.8 Flash Google | 79.1% | 61.6 | $3.75 | |
| 19 | Kimi K3 Moonshot AI | 79.0% | 62.3 | $15.00 | |
| 20 | MiMo-V2.6-Pro Xiaomi | 76.7% | 65.4 | $0.87 |
How this list is built
Generated from the dataset on every refresh. Nobody picks the order by hand.
Models by their LiveBench research result, as the board publishes it, best first. Open-ended research briefs that produce a written report, graded against a rubric. Each score is the model's mean task score for the category, 0 to 100, from a single run of the current release, shown as a share.
- OpenRouter, fetched
- LiveBench, fetched , board updated October 14, 2026