I think the reason the Claude models in Copilot seem dumb isn't just because they have a smaller context size, there's probably something else going on too. I swear on it, but I can't prove it.
In Windsurf, the exact same model performs much better even before reaching that context limit.