Anthropic's latest Claude model can work for 30 hours on its own
Add Axios as your preferred source to
see more of our stories on Google.
Image: Anthropic
Claude Sonnet 4.5, released Monday, outperforms prior versions at coding, finance, cybersecurity and long-duration autonomous work, Anthropic said.
Why it matters: To act as an agent, AI models must sustain work on a single task for hours — something many earlier models couldn't do.
Driving the news: The new version of Claude can work for 30 hours or more on its own, a big step up from the seven hours of autonomous work with Claude Opus 4.
- Beyond math and coding, where Claude has previously excelled, Sonnet 4.5 is strong on tasks requiring research and diligence, Scott White, a product lead at Anthropic, told Axios.
- The company offered a variety of benchmarks and customer comments touting the power and performance of the new model, which is priced the same as its predecessor, Claude Sonnet 4.0.
- Anthropic will give developers access to Claude Code's building blocks — virtual machines, memory and context management — to make it easier to create Claude-powered agents.
The big picture: Anthropic said the rapid progress, marked by major Sonnet updates in February and May, shows a pattern where every six months its new model can handle tasks that are twice as complex.
- "This is a continued evolution on Claude, going from an assistant to more of a collaborator to a full, autonomous agent that's capable of working for extended time horizons," White said.
Between the lines: Anthropic also claims that the new Sonnet is the company's "most-aligned" model yet.
- "Claude's improved capabilities and our extensive safety training have allowed us to substantially improve the model's behavior, reducing concerning behaviors like sycophancy, deception, power-seeking, and the tendency to encourage delusional thinking," the company said in a blog post.
- Anthropic also said that it has made "considerable progress" on defending against prompt injection attacks when Claude is autonomously accessing the web or conducting other "agentic" tasks.
