Anthropic’s “smartest model” is getting a big boost, the company said in a blog post announcing Cloud Opus 4.6.
It called the new model a “direct upgrade” from its predecessor in a release, noting that it can perform complex, multi-step tasks better and “get much closer to production-ready quality on the first try than we’ve seen with any model – documents, spreadsheets and presentations will require less back-and-forth over iterations.” It is available from today at the same price as its predecessor, and according to the company, its particular strengths lie in agentic coding, tool usage, search, and financial analysis.
But above all, it seems that with this release, Anthropic wants to extend the current promotion of the cloud beyond coding and corner the market on other types of knowledge work. With Opus 4.6, it invested in improving the models, creating presentations in PowerPoint and documents in Excel. The blog post included a plug for Cowork, Anthropic’s recent release which is a non-tech-worker-friendly version of Cloud Code, in the hopes that users in non-tech industries will explore use cases for research, marketing, and more.
On the coding front, Anthropic said in a release that Opus 4.6 was built to even better developers’ experience with cloud code, as it specializes in long-horizon tasks and “can take a development project that would normally take several days and complete it in hours, handling everything from architecture to deployment.”
The company has also announced a feature called “Agent Teams”, currently in research preview, which allows the new model to work within cloud code “the way a real engineering team does”, meaning it’s possible to divide the work of a project across agents who own a part of the project and coordinate with each other.
Dianne Na Penn, head of research product management at Anthropic, said: The Verge The company focused on improving the “multi-agent” experience for developers with this launch, investing in output quality and speed, as well as improving models across types of knowledge tasks other than coding – i.e., Excel, PowerPoint, and search functions.
“This is the first version of the Opus model where we have one million reference windows offered in beta,” Penn said. “We had so much positive feedback about Opus 4.5 that one of the key features people wanted was a longer context window so they could work with the cloud on more documents.”
Anthropic said in a blog post that it ran the “most comprehensive” set of security tests for Opus 4.6 of any of its models to date. New evaluations include assessments for user well-being, more complex tests on whether Opus 4.6 can reject “potentially dangerous requests”, and updated tests for how well the model can secretly perform harmful actions. According to the company, the model also features enhanced cybersecurity capabilities, so it includes six new cybersecurity checks to track potential misuse.
