Opus 5 is cheaper and smarter. The bill hides in a place you don't check.

Anthropic shipped Claude Opus 5 last Friday, and the early read is clear. It is strong, it is cheaper than the top model, and it comes with fewer of the guardrails that annoyed people. If you own an AI backend, this one earns a real look. But the good news comes with a catch that shows up on your invoice, not in your output. Let me catch you up.

Frontier work at half the price

The pitch is simple. Opus 5 gets close to Anthropic's most capable model, Fable 5, for half the cost. Pricing lands at $5 per million input tokens and $25 per million output, the same as the model it replaces. So you are paying last generation's price for a real jump in quality.

The benchmark story backs it up. Anthropic's own numbers say Opus 5 more than doubles the prior model's score on one coding test, and on a computer-use test it beats Fable's best result at about a third of the cost. Early customers reported concrete wins too: one legal team saw first-turn redline scores nearly double, and a trading firm hit its answers using roughly a seventh of the reasoning tokens.

For your roadmap, that means the cost math for putting a strong model in front of users just got easier. You do not have to reach for the priciest tier to get near-top output.

The rules got lighter

Opus 5 drops a lot of friction. TechCrunch reports that Anthropic expects its safety classifiers to fire 85% less often on Opus 5 than on Fable. It also skips the 30-day data retention policy that made privacy-minded users nervous. Fewer false refusals, fewer surprise blocks.

Anthropic paired that with a new beta called Automatic Fallbacks. When a prompt trips a safety filter, the system routes it to a weaker model instead of throwing an error. So an API call that used to die now returns something usable. Worth turning on if broken calls are hurting your users.

There are still hard limits, mostly around cybersecurity. Opus 5 can look for bugs in source code but is blocked from scanning software binaries for holes. Know where those walls sit before you promise a feature that leans on them.

Brilliant, and kind of exhausting

Strong does not mean pleasant. Claire Vo, who ran Opus 5 through her 7-model hands-on eval, calls its personality "neurotic." In one coding session it flat refused to touch a merge conflict. She also flags a verbosity problem she calls "Claude Slop," the kind of over-explaining that makes real work slower.

The flip side of that fussiness is judgment. Early testers liked that Opus 5 checks its own work and pushes back. One customer told Anthropic the model argued against a proposed design and "didn't fold when I insisted," then offered a compromise that kept the good part. That is useful when you want a careful reviewer. It is a drag when you want a fast answer.

Which means you cannot just swap Opus 5 in everywhere. Match it to the job. Careful, high-stakes work rewards its caution. Quick, low-stakes tasks pay the verbosity tax for nothing.

The dial nobody tells you to change

Here is the part that will bite you if you skip it. Opus 5 lets you set an effort level, low to max, and that setting moves both quality and cost hard. Anthropic's own charts show performance climbing with effort, and thinking gets billed as output tokens.

The risk shows up most with Fable in Claude Code. As Anup Karanjkar warns, that model "cannot stop thinking, and thinking is billed as output." On the default setting it bills you to deliberate over a simple git status check. The cost does not show up in worse output. It shows up in a line on the invoice you never open. Opus 5 gives you the effort control Fable does not, but only if someone sets it on purpose.

The deep cut

The money story is not the sticker price. It is the default effort setting nobody on your team touched. A cheaper model run at max effort on routine tasks can cost you more than a pricier model tuned right. Before you roll Opus 5 out, pick an effort level per use case and write it down. High effort for the vague, hard, high-stakes jobs where its judgment pays off. Low effort for the fast, cheap, boring ones. Ethan Mollick's advice on picking a thinking level points the same way: the model and the effort are two separate choices, and skipping the second one is how bills balloon. This is a config decision, and right now it is probably nobody's job.

Three questions for your team

1. What effort level are we running per use case, and who signed off on it? If the answer is "the default," that is your first bill review item.
2. Where does Opus 5's caution help us, and where does it just slow users down? Route careful work to it and keep a faster path for the low-stakes stuff.
3. Do our features touch anything Opus 5 still blocks, like binary scanning, and have we turned on Automatic Fallbacks so filtered calls return a usable answer instead of an error?