Anthropic launched Claude Sonnet 5, an improved and more agentic version of its mid-sized model. According to the company, this release marks an important step as it allows for the execution of complex tasks autonomously, something that until recently was only achieved by larger and more expensive models.
The model can plan, use tools like browsers and terminals, and carry out tasks without constant human intervention. This aligns with what other companies in the sector are doing, where agentic capabilities are already a minimum standard at all price levels.
What now differentiates the models is not just who does the agentic work better, but who does it cheaper and more reliably without human supervision.
Better performance at a lower cost
Sonnet 5 promises performance close to that of Opus 4.8, but at much lower costs. As of this Tuesday, it became the default model for free and Pro plans, and is available for all subscriptions.
At launch, the price is $2 per million input tokens and $10 per million output tokens until August 31. After that, it rises to $3 and $15 respectively. This makes it more economical than Opus 4.8, OpenAI's GPT-5.5, and Google's Gemini 3.1 Pro.
Compared to its predecessor Sonnet 4.6, launched in February, it shows clear improvements in reasoning, tool usage, coding, and knowledge tasks. In an agentic coding benchmark, for example, it achieved 63.2%, surpassing the 58.1% of the previous model and approaching the 69.2% of Opus.
In knowledge tasks, it even slightly outperforms Opus 4.8 in some cases. “Opus remains ideal for greater accuracy in complex problems, but Sonnet 5 offers developers cheaper and higher-quality options than before,” stated Anthropic.
Improvements in complex tasks and security
Testers highlighted that the model completes complex tasks where previous versions would stop. Additionally, it verifies its own output without being explicitly asked. A Zapier engineer reported that he assigned it a two-part task—updating account tiers in Salesforce and sending an announcement to business contacts—and the model completed it from start to finish.
In terms of security, Sonnet 5 reduces unwanted behaviors such as cooperation with misuse or deception. It is better at rejecting malicious requests and avoiding prompt injection attacks. It also hallucinates less and shows less sycophantic behavior than its predecessor.
Although it does not reach the level of Opus 4.8 in some aspects of alignment, its evaluations show lower capability for dangerous cybersecurity tasks. Fabian Hedin from Lovable noted that the model cleanly and consistently rejects unsafe requests.
