LiveBench 2.0 Breaks Down the Frontier: Fable 5.1 Hits Perfect 100 on Browser Use, Gemini 4 Argon Leads Reasoning at 95.1
LiveBench 2.0 publishes capability scores across seven distinct columns — Agentic, Reasoning, Coding, Browser Use, Data Analysis, Research, and Design — and the column-by-column breakdown tells a different story from the overall rankings.
Claude Fable 5.1 sits at the top overall at 88.0, but that number masks significant dispersion. No model leads every column. The frontier has fractured along capability lines, and the data now makes explicit which model to deploy for which task.
Browser Use: Fable 5.1 at 100, Argon at 92
The Browser Use column is the sharpest differentiator in the current snapshot. Claude Fable 5.1 scores 100.0 — a perfect score on the browser tasks in the benchmark. Claude Fable 5 also scores 92.0, ahead of Gemini 4 Argon at 92.0 and GPT-6.1 Sol at 88.0. GPT-6 Astra and Claude Sonnet 5.5 both sit at 84.0. Claude Opus 5.5 trails the group at 80.0.
The gap between Fable 5.1 and Opus 5.5 on Browser Use is 20 points, despite Opus 5.5 being only 1.5 points behind overall. This suggests Anthropic’s Fable line has distinct web interaction tuning that Opus does not share, even though Opus 5.5 leads on multiple other dimensions.
Browser Use as a category tests whether a model can navigate real web interfaces, complete multi-step tasks that require following links and filling forms, and synthesize information from multiple pages — the kind of work that autonomous web agents perform in production. A 100% score signals Fable 5.1 is within measurement noise of saturating this benchmark.
Reasoning: Gemini 4 Argon at 95.1
Gemini 4 Argon leads Reasoning at 95.1, ahead of GPT-6 Astra (92.7), GPT-6.1 Sol (92.6), Claude Opus 5.5 (92.2), and Claude Fable 5.1 (91.7). The spread at the top is only 3.4 points, but Argon’s Reasoning score is notably higher than its overall (84.1) — meaning Reasoning is Argon’s strongest suit and other categories pull its composite down.
The same pattern appears in Data Analysis: Argon leads at 85.2, ahead of GPT-6 Astra (83.0), GPT-6.1 Sol (82.7), and Claude models in the 59.5–80.5 range. Claude Sonnet 5.5 in particular drops to 59.5 on Data Analysis, a notable weakness for a model that otherwise scores 82.5 overall.
Coding: Claude Fable 5 at 98.3
The Coding column is dominated by Anthropic. Claude Fable 5 leads at 98.3, Claude Fable 5.1 follows at 97.7, GPT-6.1 Sol sits at 96.0, and Claude Sonnet 5.5 at 94.7. Gemini 4 Argon trails at 89.7 — nearly 9 points behind the coding leaders despite its Reasoning strength.
For software engineering workloads, this ordering matters: the Fable line outperforms GPT-6 Astra (93.7) and holds a larger lead over Argon than any other category.
Agentic: Opus 5.5 Leads, Sol Underperforms
Claude Opus 5.5 leads the Agentic column at 85.0, followed by GPT-6 Astra (82.4) and Claude Sonnet 5.5 (78.8). GPT-6.1 Sol scores only 73.0 on Agentic — below its overall ranking position — while Gemini 4 Argon scores 69.9, the weakest in the top five on this dimension.
This is the column most relevant to autonomous agent deployments: models that complete multi-step tasks, maintain context across tool calls, and recover from errors. Opus 5.5’s lead here aligns with Anthropic’s sustained investment in extended context and tool-use reliability.
Research and Design
Research is led by Claude Opus 5 at 88.9 and Claude Fable 5 at 86.9 — both older than Fable 5.1, which scores 82.0. The newer model is not the best at long-form research synthesis, a counter-intuitive result that suggests training objective shifts between model generations.
Design is a Anthropic sweep: Opus 5.5 at 86.9, Sonnet 5.5 at 85.0, Fable 5 at 84.7, Fable 5.1 at 83.4. GPT and Gemini models cluster between 73.7 and 82.1.
The Cost Outlier
GPT-6.1 Sol scores 85.4 overall at $0.734 per successful task. Claude Fable 5.1 scores 88.0 at $4.526 per task — 6.2x more expensive for 2.6 points of overall performance. For most production workloads where cost scales with volume, Sol’s per-task price is the practical choice unless Browser Use performance or Coding headroom is the decisive factor.
The full capability matrix from LiveBench 2.0 makes model selection more tractable: match the task profile to the column leader, then apply cost discipline. No single model wins everywhere.
| Model | Overall | Agentic | Reasoning | Coding | Browser Use | Data Analysis | Design | Cost/Task |
|---|---|---|---|---|---|---|---|---|
| Claude Fable 5.1 | 88.0 | 80.8 | 91.7 | 97.7 | 100.0 | 80.3 | 83.4 | $4.526 |
| Claude Opus 5.5 | 86.5 | 85.0 | 92.2 | 96.3 | 80.0 | 80.3 | 86.9 | $4.216 |
| GPT-6 Astra | 86.1 | 82.4 | 92.7 | 93.7 | 84.0 | 83.0 | 82.0 | $3.442 |
| GPT-6.1 Sol | 85.4 | 73.0 | 92.6 | 96.0 | 88.0 | 82.7 | 82.1 | $0.734 |
| Gemini 4 Argon | 84.1 | 69.9 | 95.1 | 89.7 | 92.0 | 85.2 | 73.7 | $1.830 |
| Claude Fable 5 | 83.6 | 53.2 | 89.7 | 98.3 | 92.0 | 80.5 | 84.7 | $5.505 |
| Claude Sonnet 5.5 | 82.5 | 78.8 | 91.6 | 94.7 | 84.0 | 59.5 | 85.0 | $3.805 |
| Claude Opus 5 | 82.1 | 62.1 | 91.2 | 90.8 | 84.0 | 74.6 | 83.0 | $3.907 |