Anthropic’s new hybrid AI model can work on tasks autonomously for hours at a time

While Claude Opus 4 will be limited to paying Anthropic customers, a second model, Claude Sonnet 4, will be available for both paid and free tiers of users. Opus 4 is being marketed as a powerful, large model for complex challenges, while Sonnet 4 is described as a smart, efficient model for everyday use.

Both of the new models are hybrid, meaning they can offer a swift reply or a deeper, more reasoned response depending on the nature of a request. While they calculate a response, both models can search the web or use other tools to improve their output.

AI companies are currently locked in a race to create truly useful AI agents that are able to plan, reason, and execute complex tasks both reliably and free from human supervision, says Stefano Albrecht, director of AI at the startup DeepFlow and coauthor of Multi-Agent Reinforcement Learning: Foundations and Modern Approaches. Often this involves autonomously using the internet or other tools. There are still safety and security obstacles to overcome. AI agents powered by large language models can act erratically and perform unintended actions—which becomes even more of a problem when they’re trusted to act without human supervision.

“The more agents are able to go ahead and do something over extended periods of time, the more helpful they will be, if I have to intervene less and less,” he says. “The new models’ ability to use tools in parallel is interesting—that could save some time along the way, so that’s going to be useful.”

As an example of the sorts of safety issues AI companies are still tackling, agents can end up taking unexpected shortcuts or exploiting loopholes to reach the goals they’ve been given. For example, they might book every seat on a plane to ensure that their user gets a seat, or resort to creative cheating to win a chess game. Anthropic says it managed to reduce this behavior, known as reward hacking, in both new models by 65% relative to Claude Sonnet 3.7. It achieved this by more closely monitoring problematic behaviors during training, and improving both the AI’s training environment and the evaluation methods.