Anthropic unveils Claude Sonnet 4.6, its new reference AI.

© Anthropic

As explained by Anthropic in its statement, Sonnet 4.6 is programmed to offer high-performance capabilities close to those of Opus 4.5, while remaining accessible. The model stands out by maintaining a contained token consumption. Anthropic indicates that the pricing for Claude Sonnet 4.6 remains the same as for previous Sonnet versions, at 3 dollars per million tokens sent and 15 dollars per million tokens generated by the AI.

« Performances that would have previously required the use of an Opus-class model — including for concrete and important office tasks — are now available with Sonnet 4.6 », explains Anthropic in a statement.

Also read: « Claude will remain ad-free » – Anthropic mocks ads in ChatGPT

Improvements in Claude Sonnet 4.6

On coding benchmarks and tasks related to AI agent work, Sonnet 4.6 achieves very high scores, often comparable to Opus for complex software development tasks, while making fewer false success claims. Anthropic also notes that 70% of developers prefer to converse with Sonnet 4.6 over Sonnet 4.5. Nearly 60% of developers even prefer it over Opus 4.5 for instruction following and reliability during long sessions.

According to testers, the new version of Claude is less likely to complicate its responses. Moreover, the AI is more likely to complete a complicated task through. Anthropic reveals that Claude Sonnet 4.6 crosses a major milestone in the ability of AI to control a computer autonomously. In sixteen months, Sonnet models have gone from 14.9% to 72.5% of tasks successfully completed in perfect autonomy, without any human intervention. Specifically, Sonnet 4.6 can handle complex spreadsheets, fill out multi-step online forms, and navigate between different tabs of a browser.

It is one of the most performing models for office task automation. The model has not yet surpassed human experts, but it is progressing enough for computer automation to become truly useful on many daily tasks, notes the American startup.

In the wake of this, Anthropic has improved Claude's resistance to attacks by injecting hidden queries into phishing websites. These hidden queries are designed to manipulate the AI and force it to perform malicious actions, such as data theft. According to Anthropic, this iteration of Claude is as resistant as Opus and offers a major improvement over Sonnet 4.5. In other words, cyberattacks are less likely to succeed. In summary, Sonnet 4.6 enjoys a generally warm, honest, prosocial and sometimes funny character, very robust security behaviors and no signs of major problems related to severe forms of misalignment.