I do think that Claude’s sonnet is a very good AI to use in the agent and assistant framework. However, sometimes I wish I could use 01, and as 03 looms near release, I definitely wish I could hook that up to the agent.
At some point in the past, we did have the ability to choose which large language model was being used to help generate code, but for some reason, that no longer exists. Is there any plan to bring it back?
You can select between openai/anthropic for the Assistant, for the Agent it got disabled because the reliability of agent’s actions at different steps goes down enormously when you don’t use specific models for certain actions, and we use a bunch of little models for the sub-agent actions to speed things up and keep costs down (for you and for Replit, e.g. doing a full agent run with o1 or even sonnet-3.5 only would be way more expensive vs using the ensemble of models at different steps)