Opus 4.7 just dropped, and in this video I’m breaking down what actually happened—because the real story starts with Opus 4.6. Over the last couple weeks, a ton of power users (including me) felt the model getting noticeably worse: shallow thinking, skipping file reads in Claude Code, hallucinated details like fake commit hashes and package names, and even abandoning tasks mid-way. A big part of that came down to behavior changes like adaptive thinking allocating zero reasoning tokens on “simple” turns, plus effort defaults quietly getting dropped to medium—so people were paying the same price for a worse product and burning through tokens faster when things went off the rails. Then Anthropic moves fast and releases Opus 4.7 with claims that basically map 1:1 to the complaints: better instruction following, more rigorous long-running work, self-verification, better vision, and a new X high effort mode. I walk through the benchmarks, what’s likely real model change (vision gains, tokenizer updates, SWBench Pro jump), and what’s still “take it with a grain of salt.” I also rant a bit about the desktop app launch—cool concept, but way too many obvious bugs for a company this big. Bottom line: 4.7 looks better on paper, but you still need to test it in your own Claude Code workflows before you declare it a win.