I just gave Claude Code a hideous coffee shop website and told it to make it better 100 times in a row. I ran it twice, once with Sonnet 5.5 and once with Opus 5.5, both on medium. Every round was a fresh session with no skills, CLAUDE.md, plugins, or memory. The prompt was the same every time: look at the site, innovate, make one bold change, use any tool you have, and ship something every round. What happened: - Sonnet played it safe and spent 100 rounds making small tweaks to the ordering flow. It finished in under 2 hours for about $22 in API value. - Opus kept inventing (voice ordering, a chat assistant, a pumpkin you can carve, a Spanish version) until the site hit 21,000 lines and started breaking its own features. It took 8.5 hours and about $137. - Neither model searched the web once in 200 rounds, even though the tool was right there. If you want AI to go find inspiration, you have to ask it to.