The Anthropic company presented the new Claude Sonnet 5 model of artificial intelligence, during the development of which the main attention was focused on the effective management of AI agents.
According to the company, Claude Sonnet 5 is able to independently schedule actions, use a browser and terminal, and perform complex tasks, working almost completely autonomously. Anthropic notes that until recently such capabilities were characteristic only of more powerful and expensive models.
The Claude Sonnet 5 is considered a mid-range model, but offers performance similar to the Claude Opus 4.8 at a significantly lower price. The new model has already become the default version of the Claude platform for both free and paid subscribers.
The cost of using Sonnet 5 through the API is $2 for 1 million incoming tokens and $10 for 1 million outgoing tokens until August 31st. After that, the tariffs will increase, reaching 3 and 15 dollars, respectively.
According to data provided by Anthropic, in one of the programming tests, Claude Sonnet 5 scored 63.2%, while Opus 4.8 scored 69.2% and the previous Sonnet 4.6 scored 58.1%. In some data processing tests, Sonnet 5 even outperformed Opus 4.8.
The company notes that Opus 4.8 remains the preferred choice in cases where maximum accuracy is required, but Sonnet 5 offers developers a significantly more affordable solution with high efficiency and a balance of cost and performance.
The new model has also become safer. Sonnet 5 exhibits less "unwanted behavior", is more resistant to malicious commands and query attacks, and hallucinates less often than the previous version, and gives less user-friendly responses.
Meanwhile, Anthropic notes that Sonnet 5 still lags behind Claude Opus 4.8 and Claude Mythos Preview when it comes to solving the most complex problems. According to the company, Sonnet 5 also has significantly more limited capabilities in terms of performing dangerous operations in the field of cyber security, which is considered as an additional security advantage.