alt Hacker NewsAdditionally, if there’s more “knowledge” of the code and best practices and such baked into your model, it has to spend fewer tokens reasoning about it or even potentially looking it up. I think even one web search to retrieve additional information not also in the model will probably blow up any supposed benefit you’d get by a slightly more token-efficient language.
That could be mitigated by using a much smaller non-reasoning model for searches and reading. There's no reason why a search should take loads of tokens.