What's new

BigCodeBench

coding

BigCodeBench evaluates practical code generation with 1,140 challenging tasks that require using libraries like NumPy, Pandas, and Matplotlib — going beyond simple algorithmic puzzles to test real-world software development.

View paper / source

5

Models Tested

74.0

Best Score

70.4

Average Score

0–100

Scale Range

1.1x

Weight

How It Works

Models must generate complete Python functions that use complex library APIs correctly. Each solution is tested against comprehensive unit tests. Tasks involve data processing, visualisation, file I/O, and multi-library integration.

Why It Matters

Real programming involves using libraries and frameworks, not just writing algorithms from scratch. BigCodeBench tests whether models can write the kind of code that developers actually write every day.

Limitations

Python-only. Library versions and API changes can affect results over time. Some tasks may be solvable through pattern matching of common library usage patterns.

Leaderboard — BigCodeBench

# Model Provider Score
🥇 o3 OpenAI 74.0
🥈 GPT-5.2 OpenAI 73.0
🥉 Claude Opus 4.6 Anthropic 72.0
4 Gemini 2.5 Pro Preview 06-05 Google 68.0
5 R1 DeepSeek 65.0
All Benchmarks