Anthropic Says Its Newest AI Model Is Getting Pretty Good at Using a Computer


The best Claude AI model you can get without paying for a subscription is getting a significant  upgrade, Anthropic said Tuesday. The company released Claude Sonnet 4.6, a new version of its midrange model that it said can code about as well as a previous version of the bigger Opus.

One particular improvement Anthropic highlighted about Sonnet 4.6 is its ability to use a computer the way you might, filling out forms and switching between browser tabs. In the OSWorld benchmark, which evaluates how well an AI can use an operating system, Sonnet 4.6 has shown it can operate a computer at a human baseline level, Anthropic said. That means it doesn’t necessarily need specific software connectors or tools to do things like follow a spreadsheet or browse the internet.

AI Atlas

As AI models become more capable of doing things on our behalf rather than just giving us answers, the security risks increase. A big hazard is called prompt injection: Think of it as a website hiding a command somewhere that humans won’t notice, but an AI will. (It’s one of the major risks dogging the viral AI agent OpenClaw.)

Anthropic said in its tests, Sonnet 4.6 showed significant improvement compared to Sonnet 4.5 in resisting prompt injection attacks. It was similar to Opus 4.6, released two weeks ago and only available for paid subscribers.

As a coding model, Sonnet 4.6 can better follow detailed instructions, Anthropic said. The company is beta testing a context window of 1 million tokens for the model, which means you can give the AI massive amounts of information in a single request.

Read more: I Vibe Coded an App With 3 Popular Chatbots. The Real Winner Is a Good Prompt

Claude has seen a surge in popularity in recent months, with the Claude Code app experiencing a viral moment over the holidays as people discovered its vibe coding capabilities. Anthropic launched a Super Bowl ad campaign attacking rival OpenAI for its decision to put ads in its free and low-cost ChatGPT plans. At the same time, OpenAI’s own Codex tool and latest model, GPT-5.3-codex, has emerged in recent weeks as a capable rival of Claude Code.




Leave a Reply

Your email address will not be published. Required fields are marked *