Best Open-Weight Agentic Score on LiveBench 2.0 Is 35.7, Less Than Half of Claude Fable 5.1's 77.9
On LiveBench 2.0, the strongest open-weight model overall is DeepSeek V4.1 Flash 1017 at 72.9, against 85.7 for Claude Fable 5.1. That 12.8-point gap understates where open weights fall short. Split by category, they are level on coding and far behind on agents.
The Open-Weight Field
| Model | Overall | Agentic | Coding | Browser Use | $/task |
|---|---|---|---|---|---|
| DeepSeek V4.1 Flash 1017 | 72.9 | 35.7 | 96.0 | 67.2 | $0.185 |
| MiMo V2.6 Pro | 68.5 | 35.2 | 81.5 | 77.3 | $0.183 |
| MiniMax M3.1 Flash | 66.9 | 28.8 | 83.9 | 80.6 | $0.153 |
| Kimi K3 | 64.6 | 26.4 | 59.3 | 75.2 | $2.089 |
| Mistral Large 4 1029 | 63.5 | 27.5 | 95.4 | 17.4 | $0.487 |
| GLM-5.3 Flash | 62.1 | 19.0 | 80.6 | 59.8 | $0.163 |
| Qwen3.8 Max | 61.6 | 29.9 | 60.3 | 49.4 | $1.880 |
Kimi K3.1 Pro (71.3 overall, 39.1 Agentic) scores higher than every row above except DeepSeek, but its weights are scheduled for November 4 and it is not tagged open on the board yet. If they ship as planned, it would take the open-weight Agentic lead by 3.4 points.
Where the Gap Is
- Coding: DeepSeek V4.1 Flash 1017 scores 96.0, identical to GPT-6.1 Sol. Mistral Large 4 1029 scores 95.4. Only Fable 5.1 (97.7) and Opus 5.5 (96.3) are higher.
- Agentic: the best open score is 35.7. Fable 5.1 scores 77.9, GPT-6 Astra 77.5, Opus 5.5 72.8, Sol 68.7. Even GPT-6.1 Terra, the cheapest closed model in the top seven, scores 48.6.
- Browser Use: MiniMax M3.1 Flash reaches 80.6 and MiMo V2.6 Pro 77.3, within 13 to 16 points of Sol (93.3). Mistral Large 4 1029 sits at 17.4 in the same category.
LiveBench weights the Agentic suite at 2x in its overall score, which explains most of the overall deficit.
Cost Is the Counterweight
DeepSeek V4.1 Flash 1017 costs $0.185 per successful task. Fable 5.1 costs $7.406, roughly 40 times more, for 12.8 more overall points. GPT-6.1 Terra at $0.665 and Sol at $0.819 are the closest closed-model comparisons, and cost roughly four times what DeepSeek does.
The practical split for builders is clear in the data. If the job is generating code or retrieving and summarising information at volume, open weights deliver frontier-level scores at a fraction of closed pricing. If the job is a multi-step agent that tracks state and recovers from tool failures, no open-weight model on the board scores above 36.