Anthropic launches Sonnet 5, its new AI model for maximum efficiency of agents and daily tasks, at a lower cost

Anthropic has launched Sonnet5, a new version of its Artificial Intelligence (AI) model that It is not only dedicated to answering daily queries of people, but also to the management of AI agents, with more autonomous capabilities and at a lower cost than other models.

The new Sonnet 5 model allows from using browsers and terminals to functioning autonomouslyand takes the same nomenclature as Fable 5, Anthropic’s Mythos-class model, whose deployment was paused a few days after being launched two weeks ago and which is now available again globally, after approval by the United States Government.

The AI ​​laboratory has published all the details on its website, where it has stressed that “just a few months ago” offering a performance similar to what they have achieved with this new model, which is even close to the recent models of the Opus class, would have required a larger and more expensive model.

Sonnet 5, compared to Sonnet 4.6, has a lower rate of undesirable behaviors and, generally speaking, it is safer in contexts with agents. It does have a lesser capacity for cybersecurity tasks, where Fable 5, as a Mythos class model, totally excels.

In performance tests, Sonnet 5 outperforms Sonnet 4.6 in all published tests and The most striking advance is seen in the ability to complete complex tasks end-to-end which demonstrates its high capacity for autonomy, as in Humanity’s Last Exam (expert knowledge) and Terminal-Bench v2.1 (terminal tasks) with increases of 10.6 and 13.4 points respectively in these last two.

When comparing Sonnet 5 with Opus 4.8, This mid-range model is capable of challenging one of Anthropic’s heavyweights in the agentic knowledge work tests (AA-Briefcase and GDPval-AA) and even pass them. Likewise, in the Maximum Effort test (OSWorld-Verified and BrowseComp), when activating the extra high level of reasoning, it approaches Opus 4.8 at its medium-high level.

It is worth remembering that Opus 4.8 is still better at math and reasoning purewith a notable difference compared to Sonnet 5, as well as in offensive cybersecurity, where the design of the model itself limits the capabilities of Sonnet 5.

All in all, the new version of Sonnet It is available for all subscriptions as the default model for the free and Pro version of Claude, although it will have a base cost of $3 per million input tokens and 15 dollars per million output tokens.

It should be noted that Sonnet 5 uses a new ‘tokenizer’ that processes text differently to make the AI ​​smarter. However, there is a but, the same text can take up between 1 and 1.35 times more ‘tokens’. This means that an increase of up to 35 percent in token volume can be applied for the same prompt.