3 hours ago
1
Claude Sonnet 4.5 is out today and brings major coding improvements, including checkpoints, code execution, file creation and a refreshed terminal to the AI model, Anthropic said in a press release on Monday.
Claude Code gains a much-requested feature with the addition of checkpoints, allowing coders to save their progress or roll back to a previous state. Claude can now execute code and create files, such as spreadsheets, slides and documents.
On the agent side, the Claude API lets agents run longer and handle more complex tasks. With the Claude Agent SDK, developers can create their own AI agents that can better manage memory, handle permissions and work with subagents to solve tasks.
"Our Sonnet models power so much of the AI economy, whether underpinning enterprise products or as the infrastructure behind a swarm of startups," said Mike Krieger, chief people officer at Anthropic, in a statement. "Every now and then we launch a model that we believe is going to inspire even more creation. Sonnet 4.5 is one of those. We can't wait to see what people build."
Don't miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.
Claude Sonnet 4.5 is the "most aligned frontier model we've ever released," according to Anthropic. This means that Sonnet 4.5 has seen major improvements in "sycophancy, deception, power-seeking and the tendency to encourage delusional thinking." Anthropic says it's also made "considerable progress" in defending against prompt injection attacks, when bad actors use specially crafted language to trick a model into doing things it wasn't meant to do.
"Claude Sonnet 4.5 resets our expectations -- it handles 30+ hours of autonomous coding, freeing our engineers to tackle months of complex architectural work in dramatically less time while maintaining coherence across massive codebases," Sean Ward, CEO of iGent AI, said in a press release.
Claude Sonnet 4.5 comes as the AI race is heating up. While much attention has been given to OpenAI's ChatGPT and Google's Gemini, players like Anthropic have also been pushing AI technology forward.
Fans appreciate Claude for its coding ability and the chatbot's conversational nature. In GDPval, a benchmarking tool made by OpenAI, Claude Opus 4.1 was the most performant model, beating GPT-5. It could be why OpenAI was caught using Claude Code and subsequently had its access removed for violating Anthropic's terms of service. OpenAI responded by saying it's standard in the industry to evaluate competing models for accuracy and safety and that its API would still be made available to Anthropic. In August, the two companies announced the results of a joint exercise in which each company evaluated the other's models.
(Disclosure: Ziff Davis, CNET's parent company, in April filed a lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)
As Anthropic continues to excel in certain areas, it's raising billions in capital. Recently, Anthropic completed its series F fundraising round of $13 billion and is now valued at $183 billion. Anthropic also settled a $1.5 billion lawsuit with authors for illegally pirating their work earlier this month.
OSWorld, a tool that tests how AI models perform in real-world computer tasks, benchmarked Sonnet 4.5 at 61.4%, whereas Sonnet 4 was 42.2% four months prior. The Claude for Chrome extension, which is currently available to those who signed up for the waitlist last month, takes advantage of Sonnet 4.5's agentic improvements.
