Anthropic shipped Claude Opus 4.7 yesterday. Same price as 4.6. 13% lift on their internal coding benchmark. Vision input up to 3.75 megapixels. Stronger long-context reasoning. The headline most people will pull is the coding score. The line buried in the announcement is the one that matters: "Substantially improved instruction adherence, which means existing prompts may need retuning." Translation. The model now does what you actually said, not what you sloppily meant. Every lazy prompt you've been getting away with for six months is about to behave differently. This is the upgrade test. And it sorts everyone into two camps. Camp 1. Workflow lives in a chat window You upgrade. The model stops filling in your gaps. Outputs drift. You spend a week re-tuning prompts you can barely remember writing. The "AI got worse" tweets start. They didn't. Your contract was vague and the new model stopped guessing. Camp 2. Workflow lives in a folder You upgrade. The model reads the same _config/ files. The same stage briefs. The same compliance checks. The constraints were already explicit because the system needed them to be portable. The new model is more obedient, so it follows them harder. Outputs get tighter, not wronger. That's the whole difference. Not which model you use. Whether your instructions are written down somewhere a model upgrade can land cleanly on top of. What I actually changed today Nothing in the workflow. I swapped the model name in one config file and re-ran the same content pipeline I ran last week. Carousel mining, draft, render, QC. All four stages produced cleaner output because the constraints in _config/voice.md and _config/compliance.md were already specific. The model just executed them more literally. The new self-verification behaviour is the other quiet one. Opus 4.7 checks its own outputs before reporting back. That changes how I write stage acceptance criteria. Less "produce X." More "produce X and confirm Y, Z, and W are true."