OpenAI and Anthropic spent almost all of 2025 running RL to improve the coding abilities of their models - which involves running thousands of VMs that execute generated code to see if it works.

That's why the code you get from the post-November models is so much better than older models.