Anthropic shipped Claude Opus 5 today. I use these models every day, in the sessions I run and in the things I build with them, so the question I ask first is not “how good is it” but “what do I do differently tomorrow”.
This time that has a shorter answer than usual. And a more concrete one.
The core in three sentences
Opus 5 comes close to the intelligence of Claude Fable 5, the heaviest model in the family, at half the price. It costs exactly the same as its predecessor Opus 4.8: five dollars per million input tokens, twenty-five per million out. And as of today it is the default model on Claude Max and the strongest model on Claude Pro.
That last sentence is the one that matters most to most people. If you use Claude through a subscription, there is nothing to do. You are already working with it.
What it means for you
There are roughly three ways this news reaches you.
You just use Claude, through the app or Claude Code. The switch has already happened. What you will notice is shorter answers to simple questions and longer stretches of work on hard ones, and that it finishes a job outright more often instead of stopping halfway with a proposal. One early customer describes it building its own test harness when there was no live data to validate against. That is new behaviour, not a better version of old behaviour.
You build with it, through the API. Then there is more to do than swapping a model name. Thinking is on by default now, which affects both your cost and your max_tokens. And the effort setting, the dial that controls how hard the model reasons, behaves differently than you are used to. Part four of this series covers that.
You buy it, or you own the budget. Then this is the line that counts: same price per token, fewer tokens per job. One Anthropic customer measures nine percentage points higher accuracy with a third fewer steps and sixty percent less time. Another gets comparable quality using twenty-six percent fewer tokens than Opus 4.8. The rate per token does not change. The number of tokens per finished task does.
What you can already do differently today
Three things, in increasing order of effort.
Give the model a bigger job than you used to. Opus 5 is strongest on work you would normally break into pieces: several files at once, a renovation rather than a repair. Anthropic’s own guidance is explicit about this: hand over the full task up front and let it run, rather than feeding it in slices.
Turn the effort dial lower than instinct tells you. The low and medium settings now produce quality that used to require the top settings. If you have high or xhigh sitting there because Opus 4.8 needed it, you are leaving money on the table.
Delete your own verification instructions. If your prompts tell the model to double-check its work, that line can go. Opus 5 does it unprompted, and leaving the instruction in causes duplicate work. That feels counterintuitive, because “ask the model to check itself” has been sound advice for years.
What I am not immediately enthusiastic about
Two things.
The numbers come from Anthropic. That is normal for a launch, and the table they publish is more honest than most, because it includes rows where Opus 5 loses. But it remains a vendor measuring its own product, on benchmarks partly supplied by partners. Part three of this series digs into that, including the rows it does not win.
And there is a dependency here that keeps growing. The better a model finishes long jobs on its own, the less you watch what it actually does. That is the gain and the risk in one motion. With Opus 5 it stands out more, because the model stretches its own scope: it sometimes does slightly more than you asked. Useful when it guesses right, annoying when you wanted to know exactly what changed.
The rest of this series
This is part one of four. I am writing them back to back this week, because the news is fresh now and the details matter.
Part two is about effort, the dial this release actually turns on. Part three is about the benchmarks, including where Opus 5 loses to the competition. Part four is about what breaks when you build on the API and switch over.
The announcement lives at Anthropic.
Dit stuk verscheen ook in het Nederlands: Claude Opus 5 is er: wat verandert er vandaag voor jou.