Anthropic’s latest model, Opus 5.5, was released on Tuesday and establishes a new state-of-the-art in coding and knowledge work performance, according to the company. Opus is the most capable and expensive tier of Anthropic’s three-tier Claude lineup. Sonnet is somewhere in between, and Haiku is the fastest and cheapest. In particular, Anthropic said this release outperformed the larger Fable model in many benchmarks and succeeded in many unofficial tasks that Fable was unable to complete.
The new model is significantly cheaper than its predecessor. Output token fees will be $20 per million tokens in Opus 5.5 (versus $25 in the previous model). There are similar price declines in other indicators. Model execution is also faster, reflecting the overall reduction in compute required to run the model.
The new version also includes significant changes to the way Opus communicates, with Opus 5.5 being less likely to use jargon and more likely to place important information at the beginning of messages.
This release comes just two months after the release of Opus 5 on July 24th. The next tiers in the lineup, Sonnet 5.5 and Haiku 5.5, will be released “in the coming weeks” with similar performance improvements, according to the announcement.
Anthropic says Opus 5.5 is comparable to Mythos in its biology and cybersecurity features, so its release will be subject to the same safeguards as the company’s Fable models. These safeguards limit the extent to which the model can be used to discover exploits in compiled programs, develop recognizable biological weapons, and so on.
Opus 5.5 is Anthropic’s first model release since CEO Dario Amodei accepted the call to set the pace for Frontier, intentionally slowing the advancement of AI capabilities to match the rate of advancement of coordination.
“We believe that more prudence is needed to fully address risk. We need not only to invest in risk prevention, but also to adjust the rate of progress in capabilities so that risk prevention can catch up,” Amodei wrote in a post earlier this month.
Safety training for Opus 5.5 was largely similar to its predecessors, including coordination testing and pre-release evaluation by external organizations such as METR and Frontier Design. However, Anthropic stressed that more advanced training and evaluation systems are already in place for future models, including improved security and surveillance systems.
“As AI capabilities grow, public policy must play a greater role in ensuring the safety of the systems people rely on. Building that capability will take time, and we have begun putting in place the infrastructure to support it,” the blog post said. “We look forward to sharing more details about these initiatives in the near future.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
