Hacker News Reader

Instead of burning tokens on SOTA models, why not use a dirt-cheap specialised model for file editing?

Where the SOTA model just makes a cheaper model to make edits, and it does so.

Yeah I also believe that there are plenty of efficiency gains available by using different models for different tasks. Reasoning models such as opus should only be used for the main planning and decision flows, but sub operations (exploring, applying edits etc etc) could be delegated to smaller and cheaper models. You also end up with a much smaller context for the main big model

Reply on news.ycombinator.com