Ryan

I switched to Opus 4.7 the day it dropped. Switched back to 4.6 two days later. The people calling it a regression? They’re right. But most of them stop at “it got dumber” and move on. The problem runs deeper than output quality. Anthropic shipped three changes in one release. Nobody’s connecting them. The tokenizer is new. Same prompt, same content, but 4.7 chews through 1.35 to 1.47x more tokens than 4.6 did on the exact same input. Simon Willison ran the numbers on the system prompt itself and got 1.46x. Price per token didn’t change. But your bill did. Roughly 40% more expensive per session and nobody adjusted the sticker. Then there’s adaptive reasoning. Sounds great until you realize what it means. The model now decides how hard to think about your question. Not you. The model. So when people keep saying “4.7 gives me shallow answers on complex problems” … yeah. It decided your problem wasn’t worth the compute. And the part almost nobody mentions. Switching from 4.6 to 4.7 nukes your entire prompt cache. Every cached prefix, gone. Your next cold start costs 1.45x what it used to. That hits immediately. Now stack all three. You’re paying more per token. Getting less thinking per request. And your cache is wiped on day one. Cheaper inference for Anthropic. Worse economics for you. Packaged as an upgrade. The Pro subscribers burning through their limit in 3 messages aren’t doing anything wrong. They’re just the first ones feeling it. If you’re running anything real on Claude right now, stay on 4.6. It’s more predictable, more reliable, and after this tokenizer change, genuinely cheaper per task you actually complete.