Anthropic releases Opus 5.5 with lower prices and Fable-level performance

Anthropic’s newest model, Opus 5.5, was released connected Tuesday, mounting a caller state-of-the-art successful coding and cognition enactment performance. Notably, the merchandise outpaces the larger Fable exemplary successful galore benchmarks, and succeeded successful a fig of informal tasks that Fable failed to complete.

In the announcement, Anthropic called Opus 5.5 “the strongest-performing exemplary we’ve tested to date.”

The caller exemplary is besides importantly cheaper than its predecessor. Output tokens volition beryllium charged astatine $20 per mTok for Opus 5.5, compared to $25 for the erstwhile model. Other metrics person akin terms drops. The exemplary is besides faster to run, reflecting an wide driblet successful the compute requirements.

The caller mentation besides makes important changes to however Opus communicates, with the Opus 5.5 little apt to usage jargon and much apt to enactment important accusation astatine the commencement of its messages.

The motorboat comes conscionable 2 months aft the merchandise of Opus 5 on July 24th. According to the announcement, Sonnet 5.5 and Haiku 5.5 volition beryllium released “in the coming weeks,” with akin show improvements.

Anthropic says that Opus 5.5 is comparable to Mythos successful its biology and cybersecurity capabilities, truthful its merchandise is taxable to the aforesaid safeguards arsenic the company’s Fable model. Those safeguards bounds however overmuch the models tin beryllium utilized to observe exploits successful compiled programs oregon developing recognizable biologic weapons, among different tasks.

Opus 5.5 is Anthropic’s archetypal exemplary merchandise since CEO Dario Amodei embraced calls to gait the frontier, deliberately slowing down advancement connected AI capabilities to lucifer the complaint of advancement connected alignment.

“I person go convinced that afloat addressing the risks requires adjacent much prudence,” Amodei wrote in a station earlier this month, “not conscionable investing successful hazard prevention, but pacing the complaint of capabilities advancement truthful that hazard prevention has clip to support up.”

Opus 5.5’s information grooming was broadly akin to its predecessors, with alignment investigating and pre-release valuation by extracurricular organizations similar METR and Frontier Design. But Anthropic emphasized that much precocious grooming and valuation systems were already being prepared for aboriginal models, including improved information and monitoring systems.

“As AI becomes much capable, nationalist argumentation should play a larger relation successful making definite the systems radical trust connected are safe. That capableness takes clip to build, and we’ve started to enactment the infrastructure successful spot to enactment it,” the blog station reads. “We expect to stock much details connected these efforts soon.”