PrepBench - Prep-Code Generation: leaderboard

Accuracy (%) of data-preparation code that LLM agents write from a disambiguated request and a data sample on PrepBench (2026), a benchmark of natural-language data-preparation requests over imperfect real tables. The output table must match the reference; higher is better.

Metric: Prep-code accuracy (%) from the disambiguated request with a data sample and profiling, output table compared with the reference, PrepBench natural-language data-preparation tasks, LLM agent at temperature 0.7 with Clarify (budgeted, answered by a DeepSeek-V3.2 user simulator), Profile, Code and Translate actions; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1GPT-5.1 Codex85.3#295
2Gemini 3 Flash74.8#175
3Claude Sonnet 4.570.3#261
4Kimi K2 (Thinking)69.3
5Qwen 3 235B A22B67.7#628
6GLM-4.765.7#380
7DeepSeek V3.265.4#419
8Grok Code Fast 157.5#703
9Devstral 244.1#1010
10GPT-4o39.5#751

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=prepbench-prep-code-generation · How It Works · Data refreshed daily, snapshot 2026-11-21.