Post

Open vs closed AI models: the capability gap is now 3% for most "everyday tasks". Open models are basically tied with GPT and Claude on coding, following instructions, and general knowledge. The real gap? Reasoning, long context, and agentic tasks.

This wasn't a straight line. Jan 2024: closed models led by 8%. Aug 2024: gap shrinks to 0.5%. Feb 2025: DeepSeek-R1 briefly ties the top US model. Mar 2026: gap reopens to 3.3% as reasoning models pull ahead.

Capability is a moving target, a jagged frontier, and not a settled score. So the question isn't "are open models good enough" anymore. It's: good enough for what? Know your workload and pick accordingly.