Anthropic shipped Claude Opus 5.5 this morning at $4 per million input tokens and $20 per million output — down from Opus 5’s $5 and $25 — and cut cache reads from $0.50 to $0.20, a 60% reduction. Launch coverage clocked it as roughly 30% faster than its predecessor and up to 40% cheaper on “typical workloads,” because it also burns fewer tokens per task.

Hold that sentence in your head for a moment. Two claims, two settings.

The Cheap Number and the Good Number Come From Different Runs

The $4/$20 figure is real, and it sits on Anthropic’s pricing page where anyone can check it. The 40% figure is arithmetic about tokens-per-task — and the same launch write-ups that repeat it note that it depends on your workload and has not been independently verified.

Then there’s the part buried two clicks in: Opus 5.5 defaults to medium effort. Not high.

That single default matters more than the price cut, because effort is the cost lever now. The migration documentation Anthropic published for the Opus 5 line said so outright: low and medium are no longer a compromise tier, they are the primary instrument for controlling spend. Which means the price you budget and the quality you receive are the same dial, and someone has to turn it.

That someone is you.

Anthropic Moved the Default Down and Put It in the Changelog

The prior Opus API defaulted to high effort. Opus 5.5 ships at medium. A vendor just decided, on your behalf, that your application should think less — and folded the consequences of that decision into a headline discount.

It is probably the right call. Most production traffic is retrieval, formatting, and routing, and paying frontier rates to have a model brood over a JSON schema is waste. But notice the shape of the announcement: a pricing decision, dressed as a performance story, with the settings page doing the load-bearing work.

And the migration is not clean. Launch coverage flags that Opus 5.5 breaks four Opus 5 integrations. The Opus 5 migration checklist had already taught developers that reasoning is on by default, that disabling it at the top two effort tiers returns a 400 error, and that changing tools mid-conversation is its own hazard. This is not a string swap, and the engineering hours land in the same quarter as the savings.

“We budgeted a week for the version bump and spent it on effort levels,” a platform engineer running the migration put it, in a Slack DM at 1:20 a.m. as the deploy window closed. “The model got cheaper. My team did not.”

The Benchmark With an Outside Referee Is Blank

On Terminal-Bench 4.0, Anthropic reports 66.4% for Opus 5.5 against the 57.9% OpenAI reports for GPT-6 Astra. Both numbers are vendor-reported. Neither lab is grading the other’s homework, and both are running at settings each chose for itself.

The comparison coverage also notes where the independent measurement sits: ARC Prize measured Astra at 62.71% on ARC-AGI-3 in September. Opus 5.5 has no verified ARC-AGI-3 score. Its predecessor, Opus 5, posted 30.16% in July. So the one benchmark with an outside referee has a blank beside the new model, while the self-graded ones are filled in and formatted for the announcement post.

No conspiracy is required. Independent evaluation takes weeks, and vendors ship on their own schedule. But a release optimized to be legible on launch day is a release optimized to be legible on its own terms, and the missing third-party number is the one that would have been graded at settings Anthropic doesn’t control.

Who Actually Eats the Savings

The reflex on my side of the aisle is that price competition is unimpeachable and cheaper reasoning is unambiguously good. Mostly true. But a market whose comparison axis is an internal setting is a market where you cannot comparison-shop without running your own evaluations — and the buyers who won’t run them are most of the buyers, including the ones signing three-year contracts this week.

So the savings are real and the burden is new: a cost center called “deciding how hard the model should think,” applied per route, per task, per customer, appearing on nobody’s org chart. Venture models that pencil AI gross margins as a straight line down will find the line is actually a menu. Competitors now have to publish conditional pricing too, or look expensive against a spec sheet a rival wrote. And the procurement analyst who typed $4/$20 into a forecast just volunteered her company for an experiment she didn’t design and cannot see the results of until the invoice arrives.

Anthropic didn’t lie about anything this week. It shipped a dial and let the market price it as a spec sheet. The price is real, the discount is real, and the comparison everyone is repeating is close to fiction. The gap between those things is now a line item, and nobody has a code for it yet.

Sources