Anthropic asked the industry to slow down. Its next model, Opus 5.5, is cheaper, faster and better behaved than the one before it

Anthropic released Claude Opus 5.5 on Tuesday, ten days after its chief executive, Dario Amodei, published an essay asking the industry to slow the rate at which it makes these systems more capable. The company said the new model costs 40 percent less to run than Opus 5, answers more than 30 percent faster, and scored better than any model it has tested on the audit it uses to catch its models misbehaving.

Russell Brandom at TechCrunch reported the release at 9:30 in the morning, Pacific time, and CNBC followed that afternoon, and it is the first model Anthropic has put out since Amodei's essay of 12 September, titled We Must Pace the Frontier, which was the company's own promise. In it he wrote that "fully addressing the risks requires even more prudence, not just investing in risk prevention, but pacing the rate of capabilities advancement so that risk prevention has time to keep up."

Ten days

The essay argued that the benefits of the technology were large enough to be worth unusual care, and that the answer was not to stop but to pace the work, so that the people preventing harm were never far behind the people adding capability, and within days Sam Altman of OpenAI and Elon Musk had both said they agreed with him, CNBC's CJ Haddad reported.

What came out ten days later is, on Anthropic's own account, a smaller step than a new generation and a cheaper one, since the company says Opus 5.5 does most work at the level of Claude Fable 5.1, its largest model, while needing less computing to serve, and its prices have come down with it.

The price

A million input tokens now cost four dollars where Opus 5 charged five, and a million output tokens cost twenty dollars where it charged twenty-five, according to the company's pricing page and to Gigazine, which worked the figures out in yen. Reading from the cache, which Anthropic says makes up most of the cost of the long coding sessions its customers run, has fallen from fifty cents to twenty, a cut of 60 percent.

The company also raised the five-hour usage limits on its Pro, Max and Team plans, and it has given every subscriber a single reset of that limit that they can hold on to and spend on whichever day they need it, according to the announcement and to PCWorld.

"One of the things we're continuing to innovate on is how to make that thinking, how to make the answering more efficient, so it uses less tokens depending on your effort setting," Dianne Penn, Anthropic's head of product management for research and labs, told CNBC in an interview on the day.

Who checked it

Before it was released the model was tested by two outside organisations, METR and Frontier Design, and Anthropic says that on its automated behavioural audit, which runs a model through about two thousand simulated situations by Gigazine's count, Opus 5.5 produced the best scores of any model the company has measured. It also tried to work around the boundaries it had been given about 85 percent less often than Opus 5 did, the company said, and it was harder to steer off course with instructions hidden in the material it was reading.

Because the model is about as capable as Claude Mythos 5.1 at biology and at finding weaknesses in software, Anthropic has put the same limits around it that it put around Fable 5.1, so only vetted laboratories and vetted security practitioners can apply to use the new model for that work. Nat Rubio-Licht at The Deep View reported that instead of refusing those requests outright the company routes most security tasks back to Opus 4.8 and anything flagged by its biology classifiers to Opus 5.

Less talk

The change most of the early reviews led with was not the price or the tests but the way the model writes, because Opus 5, in the words of Ben Patterson at PCWorld, had been "a notorious chatterbox," fond of finishing a job and then appending a list of loose ends and things to flag that sent you back to work. Anthropic says the new one puts the important information first and uses less jargon, and it quoted one early tester who said "it writes the way I do."

Patterson, who tried it in Claude Code for PCWorld, wrote that in his limited testing the model was far more direct and answered his questions without tacking anything on for him to chase later. Box, one of the companies that tested it early, told Anthropic that its token use fell by about a third and its response times by about 40 percent with no loss of accuracy, according to Gigazine's account of the launch material.

Claire Vo, who runs the product tool ChatPRD and reviews these models for Lenny's Newsletter, had stopped using Claude for months, "not because it got dumb, but because it got annoying," she said in a review published on Tuesday, listing "the rambling, the hedging, the preachy little disclaimers on tasks that didn't need them." After a week with Opus 5.5 across four long-running tasks, a homepage redesign and an illustration test, she titled her review I left Claude for months, Opus 5.5 is why I'm back.

What it did in a day

Anthropic's own examples are about size and time: one early tester moved a 680,000-line codebase in under a day, work the company says would have taken an engineering team weeks, and another had the model audit and fix 200,000 lines in under three hours, where Opus 5 had taken more than twenty hours and two and a half times as many tokens. When Anthropic set its models to rewrite the HAProxy load balancer from C to Rust, Opus 5.5 and Fable 5.1 both passed almost every regression test, but Opus 5.5 finished in nine and a half hours to Fable's twelve and cost 51 percent less.

Anthropic tied the plainer writing back to the essay in its own post, saying that in its own use the clearer output had made the model's work easier to follow and to check, "which is a safety benefit as well as a practical one," and that the company had built the release around the same idea, better behaviour first, with the speed and the price cut coming from needing less computing to run.

Opus 5.5 is available on all of Anthropic's platforms and in GitHub Copilot from Tuesday, and the company said Sonnet 5.5 and Haiku 5.5 would follow in the coming weeks with the same changes to speed, cost and behaviour.