Anthropic released Claude Sonnet 5 this week, and it represents a shift in how models work. Sonnet 5 is agentic. It can open a browser. It can write to disk. It can run terminal commands. It works the way you think about asking an assistant to do something without you watching every step.
The model can reason about what to do next. If you ask it to find information, it will. If you ask it to build something, it will set up the tools it needs. It is not waiting for you to tell it each action. That is the big change.
Performance is near Opus 4.8 on the reasoning work that matters. Cost is significantly lower. So Anthropic is positioning this as the model where you start moving real work into agents. Not prototypes. Not experiments. Actual production agents that run unsupervised.
We are now past the point where autonomy is a feature someone might want later. It is becoming the baseline expectation. Teams are shipping agents, not chatbots. The capability bar moved.
What gets interesting is what happens when models of this caliber are making decisions in your production environment without someone in the loop. Sonnet 5 is built for that. The guardrails are there. But the risk profile is different when the model can act, think about what to do next, and keep going. That is not a tool at that point. That is a worker.

