The Vibe Father
Roundup
Data science is the one non-web domain where the benchmarks are unusually kind to us. SWE-bench Verified — the suite everyone quotes for real-world coding skill — is built from real GitHub issues in real repositories, and those repositories are overwhelmingly Python: pandas, NumPy, scikit-learn territory. The tooling of data science is the tooling SWE-bench measures. So for data work specifically, a high SWE-bench Verified score is a genuinely strong proxy, and the top of the board is trustworthy in a way it isn't for, say, DevOps. We keep the live version at /benchmarks (VCI = SWE 40 / TB 30 / LCB 30).
But there's a domain-specific danger that no benchmark captures: in data science, a model can produce code that runs perfectly and is quietly wrong. A silent dtype coercion, a merge that drops rows, a leaked label — the notebook is green, the numbers are garbage. So the theme of this post is: pick a careful model, and never trust output you haven't validated against the data.
The data-science board, July 2026
Sorted by SWE-bench Verified, which for this domain is close to a direct read. Where a score isn't published, we say so — we don't invent numbers.
| Model | SWE-bench Verified (strong proxy here) | Notes for data work |
|---|---|---|
 Claude Fable 5 | 95.0 | Top ceiling for intricate, multi-step analysis |
| Claude Opus 4.8 | 88.6 | The careful default — elite, half of Fable's price |
| Claude Sonnet 5 | 85.2 | High-volume pipeline and transform work |
 GPT-5.5 | 80.6 | Strong all-rounder, good in notebooks and the shell |
 Gemini 3.5 Flash | 79.3 | 167 tok/s — fast exploratory iteration |
 DeepSeek V4 Pro | 77.6 | $0.435 / $0.87 per M — cheap high-volume munging |
 Qwen3.7 Max | 77.3 | 204 tok/s — fastest here, great for rapid EDA |
| Gemini 3.1 Pro | 75.6 | Puzzle strength (LCB 88.5) above its repo score |
Why the careful models win here
Data science rewards reasoning over messy, exploratory code more than almost any other domain, and it punishes carelessness invisibly. That's why Opus 4.8 is our default: at 88.6 it's elite capability with a temperament that tends to reason about edge cases — nulls, dtypes, index alignment — rather than plowing ahead. Fable 5's 95.0 earns its premium on genuinely intricate analysis: a multi-stage feature pipeline, a subtle statistical transform, a bug that spans several cells. Hand it the hard analysis, then step down for the routine work.
Gemini 3.1 Pro deserves a mention: its LiveCodeBench score of 88.5 sits above its repo-surgery number, and self-contained algorithmic problem-solving is a lot of what exploratory data work actually is. It's a real option when the task is a contained transform rather than a sprawling codebase change.
The trap that no benchmark measures: silent corruption
This is the section to internalize. Every data-science model failure that hurts is the one that doesn't throw an error:
- A
mergeorjointhat silently drops or duplicates rows because the keys weren't unique. - A dtype quietly coerced — dates read as strings, integers as floats, categories mangled.
- Data leakage: a feature computed over the whole dataset before the train/test split.
- An aggregation that looks right but excludes nulls you meant to count.
No model, however good, is exempt. The mitigation is process: assert row counts before and after joins, check dtypes explicitly, validate against a known-good sample, and read what the code actually does rather than trusting a clean run. A careful model reduces these; only validation eliminates them.
Picks by data-science workload
- Intricate analysis and pipelines. Opus 4.8 as the default; Fable 5 when the transform is genuinely hard and correctness is subtle.
- High-volume data munging. DeepSeek V4 Pro for value or Sonnet 5 — capable enough for cleaning and reshaping, cheap enough to run all day.
- Fast exploratory EDA. Gemini 3.5 Flash or Qwen3.7 Max. The speed makes the load-look-transform loop feel instant, and you're validating everything anyway.
- Contained algorithmic transforms. Gemini 3.1 Pro is a strong option given its LiveCodeBench strength.
How much do the gaps matter?
Be honest about your workload. If most of your work is standard cleaning, grouping, and plotting, the gap between a 77 and an 88 shows up rarely — a capable mid-board model plus good context engineering ships plenty. The frontier scores earn their premium on the intricate analysis where a subtle error costs you a wrong conclusion. Match the model to the difficulty, not to the leaderboard.
Our honest bottom line for data science
Because this domain sits so close to what SWE-bench measures, the board is a real ranking, not a shortlist — trust it more than you would elsewhere. Run Opus 4.8 as your careful default, Fable 5 for the hard analysis, DeepSeek V4 Pro or a fast model for the high-volume and exploratory work. But whichever you pick, the discipline that separates good data science from confidently-wrong data science is validation, not model choice. For the modeling side see the best AI for machine learning engineering, for the data layer the best AI model for SQL, and the language it all runs on in the best AI model for Python. Understand the benchmark this rests on in our SWE-bench explainer. Live numbers at /benchmarks.
