Anthropic launches Claude 4.5, touts better abilities, targets business customers

By Jeffrey Dastin and Deepa Seetharaman

(Reuters) -Anthropic unveiled the Claude 4.5 AI model on Monday, saying the newest version can code for longer uninterrupted stretches and handle finance and scientific tasks better, as the startup pushes deeper into enterprise AI.

The Alphabet and Amazon.com-backed AI startup is racing rivals to build models that can reliably operate software and complete multi-step work, key for AI agents, which can perform tasks on behalf of humans.

The Sonnet 4.5 model created a web app from scratch in internal tests, and one customer had the AI chatbot code autonomously for 30 hours, up from a seven-hour run achieved by Anthropic's earlier Claude Opus 4 for a different client, Chief Product Officer Mike Krieger said.

Anthropic is targeting power users and business customers rather than chasing a viral consumer moment, he said.

Claude 4.5 is stronger at finance and scientific reasoning and better at using computers, scoring about 60% on a benchmark that tests operating-system dexterity versus roughly 40% for prior models, the company said.

"It's a lot more visceral when you just see the model using a computer the way a person does if you're not a coder," said Chief Science Officer Jared Kaplan.

Separately on Monday, Microsoft said it would add new Microsoft 365 Copilot features powered by Anthropic models, including "Agent Mode" in Excel and Word and an "Office Agent" in Copilot chat, with PowerPoint to follow.

Microsoft last week said it would bring Anthropic's models to Microsoft 365 Copilot to diversify beyond longtime partner OpenAI.

Anthropic, founded by former OpenAI executives, has positioned Claude for workplace use with guardrails it says reduce risky outputs. The company has been marketing Claude's coding and data-analysis skills to regulated industries and teams that want models to work across multiple software tools.

Krieger said the company's focus is on sustained, reliable performance over long tasks rather than short demos.

(Reporting by Jeffrey Dastin, Deepa Seetharaman in San Francisco and Akash Sriram in Bengaluru; Editing by Anil D'Silva)