Anthropic Challenges OpenAI with Claude Sonnet 5, the Affordable Agentic AI

Anthropic Unveils Claude Sonnet 5, an Economic Agentic AI

In a context where agentic capabilities are becoming essential for AI model companies, Anthropic has recently introduced Claude Sonnet 5. This mid-sized model stands out for its increased power and advanced agentic features.

According to a blog post published by Anthropic, Claude Sonnet 5 is capable of planning, using tools such as browsers and terminals, and operating autonomously. These capabilities, which previously required larger and more expensive models, are now accessible at a reduced cost.

Competition with OpenAI and Google

Anthropic is not alone in this race for agenticity. OpenAI recently launched GPT-5.6 Sol, a model that allows users to delegate tasks to sub-agents for extended autonomous operations. Meanwhile, Google has introduced Gemini 3.5 Flash, marking a transition from a simple chatbot to an agentic tool capable of planning and executing tasks with minimal human intervention.

The introduction of Sonnet 5 by Anthropic confirms that agenticity is now a standard expectation at all price levels. The real differentiation between models will be based on the ability to offer these features at a lower cost and with increased reliability without human supervision.

Promising Performance at a Reduced Cost

Claude Sonnet 5 promises performance close to that of Opus 4.8, but at significantly lower costs. Upon its launch, Sonnet 5 is offered at $2 per million input tokens and $10 per million output tokens, with a planned increase to $3 per million input tokens after August 31. This pricing makes it more competitive than models like OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, although it remains more expensive than Gemini 3.5 Flash.

Sonnet 5 will be the default model for free and Pro plans and will be available for every subscription, making it accessible to a wide range of users.

Improvements Over Sonnet 4.6

Compared to its predecessor Sonnet 4.6, launched in February, Sonnet 5 shows significant improvements in agentic performance, particularly in reasoning, tool usage, programming, and knowledge work. In an agentic coding benchmark, Sonnet 5 achieved a score of 63.2%, surpassing Sonnet 4.6, which scored 58.1%.

In a knowledge work benchmark, Sonnet 5 competes with Opus 4.8, which is recognized for its ability to solve complex problems requiring subtle judgments and in-depth research. Anthropic emphasizes that while Opus 4.8 remains the preferred choice for increased accuracy, Sonnet 5 offers a more economical alternative without sacrificing quality.

Sonnet 5 is also capable of verifying its own output without being explicitly prompted, representing a notable advancement over previous versions.

Testimonials and Enhanced Security

Testers have reported that Sonnet 5 excels in completing complex tasks where previous versions struggled. Daniel Shepard, a senior engineer at Zapier, testified to Sonnet 5's ability to successfully perform tasks such as updating Salesforce accounts and sending launch announcements without human intervention.

In terms of security, Sonnet 5 exhibits a reduced rate of undesirable behaviors, such as cooperating with abuse or deception, compared to Sonnet 4.6. It is also more effective at refusing malicious requests and avoiding hijacking during prompt injection attacks. However, it has not yet reached the level of Opus 4.8 and Claude Mythos Preview in terms of misaligned behaviors.

Sonnet 5 hallucinates and exhibits sycophantic behavior at a lower rate than Sonnet 4.6, improving its reliability in agentic contexts.

Evaluations show that Sonnet 5 has a much lower capacity to perform dangerous cybersecurity tasks than current Opus models, which could limit its use in certain sensitive areas.

Fabian Hedin, co-founder of Lovable, emphasized the importance of a model capable of refusing dangerous requests. According to him, a model that knows how to say no is as crucial as one that knows how to build, especially in an environment where powerful tools are made available to millions of creators.