An open letter to the engineers and product designers at Anthropic.

You’ve built the best model in the world. So why does Claude Code’s WebFetch tool silently route through a weaker model to summarize web pages instead of returning the actual contents?

The user has no idea this is happening. There’s no indication in the UX — you have to press Ctrl-O mid-fetch to notice it’s a subagent under the hood. If you’re running Opus 4.6 but feeding it a Haiku summary, you’re getting maybe 40% of a web page’s meaning into your main thread.

It is enormously better to return the full page as markdown — via Firecrawl or something like it. This is what TinyFat does. I’m paying for Firecrawl and passing the results forward. They’re great.

There does seem to be a way for the main model to call fetch directly and get the actual contents back. So why isn’t that the default?

Other Posts