RealBench (RAG) - Level 2: leaderboard

RealBench (2026) asks LLMs to build a whole Python repository from natural-language requirements and UML system designs, here with retrieval-augmented generation (file by file, each conditioned on the earlier generated files it imports) on the level-2 repositories (about 850 lines of code). The score is the percentage of the original unit tests that the generated repository passes, higher is better.

Metric: Test pass rate (%), passed tests over all tests of each task, averaged over the level-2 tasks (17 repositories of about 850 lines), with retrieval-augmented generation (file by file, each conditioned on the earlier generated files it imports); RealBench repository-level Python code generation from natural-language requirements plus UML package and class diagrams (61 real repositories in four size levels), one greedy solution per task through the official APIs; higher is better. Source: arxiv.org. Saturation forecast: Around 2037. 6 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 Coder 7B Instruct9.49#3976
2GPT-4o (2024-05-13)5.73#740
3Claude Sonnet 4 (20250514)5.07#410
4DeepSeek V35.03#736
5Gemini 2.5 Flash4.54#466
6Qwen 3 235B A22B4.12#633

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=realbench-rag-level-2 · How It Works · Data refreshed daily, snapshot 2026-11-22.