Anthropic releases Claude Sonnet 4.6 – it can do everything better

According to Anthropic, the latest version of Claude Sonnet is its most powerful version yet. This applies to its capabilities in coding, computer use, reasoning, agentic tasks, knowledge work, and design. New is the context window of one million tokens.

Claude Sonnet 4.6 is initially released as a beta – as were past versions. For people using Anthropic's free chatbot, it will be the default model. Even with a Pro subscription, Claude will use Sonnet 4.6 in the future. Sonnet remains the mid-tier model in the Claude model family. Haiku is particularly fast and cost-efficient, Opus is the most powerful model for complex tasks. According to the benchmarks published by Anthropic, Sonnet 4.6 now achieves test results between Opus 4.5 and Opus 4.6. As always, practical experience may vary from benchmark results depending on the task and needs.

Sonnet is significantly more cost-effective. There are even new features for developers designed to save tokens. These include context compression, which is intended to summarize older chat histories more concisely.

Claude Opus 4.6 was only released last week. In fact, extensive tasks can quickly incur high costs. Anthropic states that Opus remains the model of choice for particularly "deep reasoning" – when coordinating multiple agents or during codebase refactoring.

In the blog post, Anthropic highlights Claude Sonnet's computer use in particular. This refers to operating regular software, almost as a human would use it – without applications being explicitly made usable for AI. In the OSWorld benchmark, which tests exactly this capability, the new Claude version fulfills 72.5 percent of the tasks, according to Anthropic. These include operating LibreOffice, Chrome, and VS Code. The previous version, Claude Sonnet 4.5, achieved 61.4 percent.

However, computer use is also particularly susceptible to attacks like prompt injections, which include hidden instructions on websites. The new version of Sonnet is said to be significantly better protected than its predecessor – attacks should be detected and, of course, circumvented if possible. However, the problem is not solved by this.

(emw)