Anthropic’s latest mannequin, Opus 5.5, was launched on Tuesday, setting a brand new state-of-the-art in coding and data work efficiency, in response to the corporate. Opus is essentially the most succesful and costly tier in Anthropic’s three-tier Claude lineup; Sonnet sits within the center, whereas Haiku is the quickest and most cost-effective. Notably, Anthropic says, the discharge outpaces the bigger Fable mannequin in lots of benchmarks and succeeded in quite a lot of casual duties that Fable failed to finish.
The brand new mannequin can also be considerably cheaper than its predecessor. Output tokens can be charged at $20 per million tokens for Opus 5.5, in comparison with $25 for the earlier mannequin. Different metrics have related worth drops. The mannequin can also be sooner to run, reflecting an general drop within the compute required to serve it.
The brand new model additionally makes important modifications to how Opus communicates, with the Opus 5.5 much less doubtless to make use of jargon and extra prone to put necessary data at first of its messages.
The launch comes simply two months after the discharge of Opus 5 on July 24. Based on the announcement, Sonnet 5.5 and Haiku 5.5, that are the following tiers within the lineup, can be launched “within the coming weeks,” with related efficiency enhancements.
Anthropic says that Opus 5.5 is corresponding to Mythos in its biology and cybersecurity capabilities, so its launch is topic to the identical safeguards as the corporate’s Fable mannequin. These safeguards restrict how a lot the fashions can be utilized to find exploits in compiled packages or creating recognizable organic weapons, amongst different duties.
Opus 5.5 is Anthropic’s first mannequin launch since CEO Dario Amodei embraced calls to tempo the frontier, intentionally slowing down progress on AI capabilities to match the speed of progress on alignment.
“I’ve change into satisfied that absolutely addressing the dangers requires much more prudence,” Amodei wrote in a submit earlier this month, “not simply investing in threat prevention, however pacing the speed of capabilities development in order that threat prevention has time to maintain up.”
Opus 5.5’s security coaching was broadly much like its predecessors, with alignment testing and pre-release analysis by outdoors organizations like METR and Frontier Design. However Anthropic emphasised that extra superior coaching and analysis programs had been already being ready for future fashions, together with improved safety and monitoring programs.
“As AI turns into extra succesful, public coverage ought to play a bigger position in ensuring the programs individuals depend on are protected. That capability takes time to construct, and we’ve began to place the infrastructure in place to help it,” the weblog submit reads. “We anticipate to share extra particulars on these efforts quickly.”
If you buy by means of hyperlinks in our articles, we could earn a small fee. This doesn’t have an effect on our editorial independence.
