- Claude Sonnet 5.0 has been released by Anthropic, which claims it is its most “agentic” model yet.
- The company stated that Sonnet 5 shows an overall lower rate of undesirable behaviors than Sonnet 4.6, making it generally safer to use in agentic contexts.
- Sonnet 5 is more effective at refusing malicious requests and resisting prompt-injection attempts.
- According to the System Card, Sonnet 5 shows clear gains over Claude Sonnet 4.6 in coding, agentic search, multimodal reasoning, and professional-task performance.
- Starting in September, Sonnet users will pay $3 per million input tokens and $15 per million output tokens, with a special rate of $2 per million inputs and $10 per million outputs through the end of August.
- While Sonnet 5 can perform routine cybersecurity tasks, it is guardrailed against generating offensive attack code.
Anthropic's Claude Sonnet 5.0 has been unveiled as the company's most 'agentic' model, showcasing significant improvements in safety and performance.12346
The company reported, “Our safety assessments found that Sonnet 5 shows an overall lower rate of undesirable behaviors than Sonnet 4.6, and is generally safer to use in agentic contexts.” This new model is particularly adept at refusing malicious requests and resisting prompt-injection attempts, marking a notable advancement in its operational capabilities.
According to the System Card, “Across a broad suite of internal and third-party benchmarks, Sonnet 5 shows clear gains over Claude Sonnet 4.6 in coding, agentic search, multimodal reasoning, and professional-task performance.” Despite not being specifically trained for cybersecurity tasks, Sonnet 5 can perform routine cybersecurity functions while being guardrailed against generating offensive attack code.
Starting in September, users will be charged $3 per million input tokens and $15 per million output tokens, although a promotional rate of $2 per million inputs and $10 per million outputs is available until the end of August.5
“Anthropic has launched Claude Sonnet 5.0, which shows a lower rate of undesirable behaviors compared to its predecessor. The model is designed to perform better in various tasks while maintaining safety in agentic contexts.”
