GPT-5.6 Sol is #1 on LiveBench and costs 63% less per successful task than Claude Fable 5
At the same maximum effort setting, Sol leads overall by 1.61 points and Agentic Coding by 18.69 points.
Event date Published
Sol pairs a narrow score lead with a large cost advantage
At the maximum effort setting, GPT-5.6 Sol scores 82.4 on LiveBench versus Claude Fable 5 at 80.79, while costing $0.59 rather than $1.57 per successful task. Sol's largest category lead is Agentic Coding, 65.61 to 46.92. At xhigh effort, Sol also leads 79.66 to 79.53 and costs $0.36 rather than $0.81.
- Maximum effort score
- 82.4 vs 80.79 GPT-5.6 Sol leads Claude Fable 5 by 1.61 points.
- Maximum effort cost
- $0.59 vs $1.57 Sol costs 62.6% less per successful task.
- Agentic Coding
- 65.61 vs 46.92 Sol leads Claude Fable 5 by 18.69 points.
- Xhigh effort score
- 79.66 vs 79.53 GPT-5.6 Sol leads Claude Fable 5 by 0.13 points.
- Xhigh effort cost
- $0.36 vs $0.81 Sol costs 56.2% less per successful task.
Sol stays left of Fable while scoring slightly higher
| Configuration | Cost per successful task (USD) | LiveBench score |
|---|---|---|
| Sol maximum | $0.59 | 82.4 |
| Fable maximum | $1.57 | 80.79 |
| Sol xhigh | $0.36 | 79.66 |
| Fable xhigh | $0.81 | 79.53 |
Gold marks GPT-5.6 Sol. The chart compares score with dollars per successful LiveBench task; lower cost is farther left and higher score is farther up.
The cost gap is larger than the score gap
The overall score gap is modest, but the cost gap is substantial and the Agentic Coding gap is much wider. This is a LiveBench-specific comparison, not a claim that either model is universally more efficient.