LiveBench 2.0 Long Agentic
Agents · 2026-10-14
LiveBench 2.0 Long Agentic suite: Long-running agentic work where early decisions must hold up through later turns, simulated months or test cases (finance and operations apps, long-run operations, systems engineering, performance optimisation); 18 tasks, each scored 0-100. LiveBench 2.0 (Abacus AI, released October 14, 2026) replaced the v1 question-answer categories with seven suites, five of them agentic; each suite is one run per model at its highest supported effort, a task the run could not measure scores 0, and a model without an input a suite needs (a text-only model on Browser) scores 0 there. Not comparable with LiveBench v1.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 85.0 |
| GPT-6 Astra | 82.4 |
| Claude Fable 5.1 | 80.8 |
| Claude Sonnet 5.5 | 78.8 |
| GPT-6.1 Sol | 73.0 |
| Gemini 4 Argon | 69.9 |
| GPT-6 Sol | 66.9 |
| Claude Opus 5 | 62.1 |
| Claude Haiku 5.5 | 57.6 |
| Claude Fable 5 | 53.2 |
| Muse Wonder | 53.2 |
| DeepSeek V4.1 Flash 1017 | 50.9 |
| Grok 4.7 | 50.1 |
| Kimi K3.1 Pro | 49.5 |
| GPT-5.6 Sol | 48.5 |
| GPT-5.6 Terra | 48.4 |
| Opus 4.8 | 48.3 |
| GPT-6 Luna | 48.1 |
| MiMo-V2.6-Pro | 47.5 |
| GPT-5.5 | 44.6 |
| Kimi K3 | 42.8 |
| Grok 4.6 | 40.7 |
| Muse Spark 1.3 | 40.4 |
| DeepSeek V4.1 Flash | 40.1 |
| Sonnet 5 | 39.0 |
| DeepSeek V4 Flash Vision Exp | 39.0 |
| Qwen3.8-Flash-Next | 37.1 |
| DeepSeek-V4-Flash-0731 | 36.6 |
| GLM-5.3 | 35.9 |
| MiniMax M3.1 Flash | 34.7 |
| GPT-5.6 Luna | 30.4 |
| Gemini 3.8 Flash | 30.1 |
| GLM-5.3 Flash | 28.4 |
| Gemini 3.7 Flash | 28.3 |
| Grok 4.5 | 27.9 |
| Qwen3.8 27B | 27.9 |
| Step 5 Preview | 27.5 |
| Gemini 3.1 Pro Preview | 23.9 |
| Sonnet 4.6 | 23.1 |
| Muse Spark 1.2 | 21.3 |