If you use Claude mainly as a chat box or coding helper, the easiest mistake after a big release is treating every Claude model like the same tool with a higher score. That is how you misread Sonnet 5 and end up blaming yourself when the output feels off.
My read is simpler: Sonnet 5 is killing prompt magic. On Sonnet 5, vague prompts are a bug, not a style choice. [C002] These launches are most worth reading not for raw strength, but for why they tighten the boundary first. The part people will actually argue about is not that the model got stronger. It is that the stronger model reads your mind less, not more.
Anthropic's Sonnet 5 prompting guide says the model is more literal, and especially at lower thinking effort it will not reliably fill in rules you forgot to state. If you care about format, scope, or tone staying stable, you have to write those constraints down.
The sharpest clue is not a benchmark. It is the migration note around Sonnet 5: non-default temperature, top_p, or top_k can return a 400. That matters because old habits that tried to steer style with randomness settings do not just drift anymore; they can get rejected. That is a loud signal that fuzzy prompt craft is losing ground to explicit written instructions. [C001]
So the wrong comparison is not just 'which Claude scored higher?' The better question is 'which Claude still fits the way I write prompts?' If you try Sonnet 5, start with the job, the output format, and the limits in plain sentences. If that changes how you compare Claude releases, share this with the person who still treats them like the same tool.
Docs only here: no live product test, no head-to-head benchmark, no user feedback.