On "llm 0.35"

Simon Willison's llm tool just got access to GPT-6 Astra, which matters less for the model itself and more because it signals how fast the CLI wrapper is staying synchronized with OpenAI's releases. If you're building on top of LLM as an interface layer, you need this kind of velocity—new models land and your abstraction layer either keeps pace or becomes a liability. The real win here is that practitioners can test against fresh models without rewriting integration code, which is how you actually scale experimentation across a team.