Opus 4.7 is here, and it's the upgrade I didn't expect to care about
Anthropic dropped Opus 4.7 this week. On paper it's a point release. The kind of thing I'd normally skim the changelog for and then forget about by lunch. I've had a day with it now and I think it's the most useful Opus update since 4.0.
The headline numbers are fine. Better coding scores, a small bump on agentic tasks. None of that is what got me. What got me is that the model finally feels like it knows when to stop.
The thing that's different
If you've used 4.5 or 4.6 for long sessions, you know the failure mode. You ask for a small fix. It gives you the small fix plus a refactor of three nearby functions plus a new test file plus some docs you didn't ask for. Helpful in theory, exhausting in practice. You end up reviewing twice as much code as you wanted to write.
4.7 doesn't do that as much. I tried it on a handful of tasks this morning, some of which I'd previously run through 4.6, and the diffs came back tighter. It still does the work. It just doesn't keep going after the work is done.
I don't have a strong theory for why. Anthropic's notes mention some training adjustments aimed at "scope discipline" and they're vague about the rest. Whatever they did, you can feel it.
What's actually new
A few things worth knowing if you're going to switch.
Longer effective context. The window is the same size, but the model holds onto information from earlier in a session more reliably. I tested this with a long debugging conversation and it was still referencing decisions we'd made an hour earlier without me having to remind it.
Tool use is faster. Not in any benchmark I've seen published, just in vibes. The gap between asking for something and watching it execute feels shorter. If you're building agents that adds up.
Better at saying it doesn't know. This is maybe my favorite small change. Ask it about a library it's unsure of and you're more likely to get "I'm not certain, let me check" than a confident-sounding hallucination. This was a real wart on 4.6.
Pricing is unchanged. Same per-token cost as 4.6. No surprises on the bill.
What didn't change
The shape of the model is the same. It's still strongest on code, writing, and reasoning. It's still not a vision-first model if that's what you need. Tool integrations are the same. Context window is the same. If 4.6 was already working for you, 4.7 isn't going to feel like a different product.
Should you switch
Probably yes, if you're already paying for Opus. The API model ID is claude-opus-4-7 and it's available wherever 4.6 was. There's nothing to migrate. Update the model string and you're done.
If you're on Sonnet because Opus felt like overkill, this doesn't really change the math. Sonnet 4.6 is still the better default for most workloads. Opus is for the work that's hard enough to justify the cost.
What I keep coming back to is the discipline change. A model that does what I asked and then quits is just more useful day-to-day than one that does what I asked and then volunteers for the next three things. It's a small shift. It's also the one I'll remember from this release.
