Claude Sonnet 5.5 Scores 59.5% on LiveBench 2.0 Data Analysis, 35 Points Below Its Own Coding Score
The October 28 LiveBench 2.0 run shows Claude Sonnet 5.5 scoring 59.5% on the Data Analysis category, the lowest result among the top six models and 35 points below its own Coding score of 94.7%. The nearest competitor on Data Analysis is Claude Fable 5.1 and Claude Opus 5.5, both at 80.3%. Gemini 4 Argon leads the field at 85.2%.
Data Analysis vs Coding Across the Top Six
| Model | Overall | Coding | Data Analysis | $/task |
|---|---|---|---|---|
| Claude Fable 5.1 | 85.7 | 97.7 | 80.3 | $6.904 |
| GPT-6 Astra | 85.6 | 93.7 | 83.0 | $4.548 |
| Claude Opus 5.5 | 85.1 | 96.3 | 80.3 | $4.291 |
| GPT-6.1 Sol | 84.1 | 96.0 | 82.7 | $0.761 |
| Claude Sonnet 5.5 | 80.0 | 94.7 | 59.5 | $3.718 |
| Gemini 4 Argon | 79.3 | 89.7 | 85.2 | $2.073 |
Sonnet 5.5 competes on Coding (94.7%), Reasoning (91.6%), and Browser Use (89.3%). The Data Analysis score stands apart: 20.8 points below the next-lowest model, 25.7 points below the category leader.
What LiveBench 2.0 Data Analysis Measures
LiveBench generates new tasks each run to prevent contamination from training data. The Data Analysis category tests statistical reasoning, tabular data interpretation, and analytical problem-solving from structured inputs. It is a distinct skill from code generation, which measures syntax production, algorithmic logic, and test-passing.
Sonnet 5.5’s Coding score indicates it can write the code to run an analysis. Its Data Analysis score suggests it struggles with the interpretation stage: reading outputs, drawing correct inferences, and summarizing from structured data.
What This Means for Agentic Pipelines
Production agentic workflows rarely separate code from analysis. A standard pattern is to generate a query or script, execute it against a data source, interpret the results, and summarize findings. A model that codes at 94.7% but analyzes at 59.5% creates a failure point at the final step.
GPT-6.1 Sol scores 82.7% on Data Analysis at $0.761 per successful task versus Sonnet 5.5’s $3.718. Gemini 4 Argon leads at 85.2% and $2.073 per task. For workloads that mix code generation with data interpretation, both models outperform Sonnet 5.5 on the dimension that matters most, at lower cost.
Anthropic’s published benchmarks for Sonnet 5.5 at launch covered Terminal-Bench 4.0 (70.6%), FrontierCode 1.1 (46.2%), and OSWorld 2.1 (80.1%). No data analysis benchmark appeared in the launch materials. LiveBench 2.0 is the first public benchmark in this survey to measure the capability directly.
The 35-point intra-model gap is the widest recorded among the top-six models on this run. It does not show up in overall scores, which average across eight categories. Procurement decisions based on overall scores alone will miss it.