Developers working on frontend performance are singling out Claude Opus as the model worth paying for when PageSpeed scores need to move. The workflow being reported: tune one page at a time, get the score to a target number, document what worked, then apply that pattern to the rest of the codebase.
It's methodical work that plays to Opus's strengths. The model has a 200k token context window - roughly the length of a short novel - which means it can hold an entire page's HTML, CSS, and JavaScript in context simultaneously while reasoning through how each piece affects loading performance. Lighter models tend to make changes that technically compile but miss the root cause: a render-blocking script, an uncompressed image, a CSS rule forcing layout recalculations.
Opus is Anthropic's most capable model and also its most expensive at $15 per million input tokens, roughly 10 times the cost of Claude Haiku. For frontend cleanup specifically, practitioners seem to be finding that quality gap justifies the cost - Opus is better at identifying which changes will actually move the needle versus which ones are just tidying. PageSpeed optimization involves dozens of interdependent factors (image formats, font loading strategy, script defer attributes, CSS specificity), and getting it right requires holding all of them in mind at once.
This is consistent with a broader pattern around model selection: use the most capable model for tasks where multiple constraints interact, use smaller models for straightforward single-step tasks. Frontend performance work - where one CSS rule can undo three other optimizations - is exactly the kind of problem where spending more per token tends to pay off.