LiveBench 2.0: Claude Fable 5.1 Scores 88.0 as GPT-6.1 Sol Matches 97% at One-Seventh the Cost
LiveBench 2.0, the updated contamination-free long-horizon benchmark, has published its current leaderboard. The top five are tightly packed within 3.9 points overall, but the cost gap between first and fourth place is 7x.
Top-5 Overall Scores
| Model | Overall | Agentic | Reasoning | Coding | Browser Use | $/task |
|---|---|---|---|---|---|---|
| Claude Fable 5.1 | 88.0 | 80.8 | 91.7 | 97.7 | 100.0 | $5.096 |
| Claude Opus 5.5 | 86.5 | 85.0 | 92.2 | 96.3 | 80.0 | $4.190 |
| GPT-6 Astra | 86.1 | 82.4 | 92.7 | 93.7 | 84.0 | $3.805 |
| GPT-6.1 Sol | 85.4 | 73.0 | 92.6 | 96.0 | 88.0 | $0.726 |
| Gemini 4 Argon | 84.1 | 69.9 | 95.1 | 89.7 | 92.0 | $1.802 |
What LiveBench 2.0 Measures
LiveBench 2.0 evaluates across eight categories: Agentic, Reasoning, Coding, Browser Use, Data Analysis, Research, and Design. The benchmark generates fresh tasks to prevent contamination from training data — a persistent problem on static benchmarks where high scores partly reflect memorization rather than capability.
The Value Case for GPT-6.1 Sol
At $0.726 per successful task, GPT-6.1 Sol delivers 85.4 overall — 97% of Claude Fable 5.1’s 88.0. The cost comparison is stark: completing the same task volume at Fable 5.1 prices costs seven times more. Sol’s weakest dimension is Agentic (73.0 vs Fable’s 80.8), but it ties Fable on Coding (96.0 vs 97.7) and leads on Browser Use (88.0 vs Fable’s 100.0… at one-seventh the cost per task).
Model Strengths by Category
Fable 5.1 leads overall and is the only model at 100.0 on Browser Use. It also tops Coding at 97.7.
Opus 5.5 records the highest Agentic score in the top five at 85.0 — 4.2 points above Fable and 2.6 above Astra, suggesting it remains the strongest for long-horizon autonomous task execution despite trailing Fable on the composite.
GPT-6 Astra leads Reasoning at 92.7, edging Opus 5.5’s 92.2.
Gemini 4 Argon has the highest Reasoning score among value-tier models at 95.1 — above Fable, Opus 5.5, and Astra — and leads Browser Use at 92.0 alongside its $1.802/task price point, which sits between Sol and Astra.
Context
LiveBench 2.0 scores are up significantly from the prior LiveBench run, reflecting both new tasks and new models in the table. Claude Opus 5.5 was not ranked on previous LiveBench versions; its debut at 86.5 places it immediately in the second slot. DeepSeek V4.1 Flash, which led the September run’s Agentic Coding dimension, does not appear in the LiveBench 2.0 top five.