Anthropic’s Claude Sonnet 5 offers cheaper, safer AI agents amid rising competition
Anthropic has released Claude Sonnet 5, a midsize model designed to handle autonomous agent tasks at a lower price point than its predecessor and several competitor offerings.
Anthropic has released Claude Sonnet 5, a new iteration of its midsize model, as the company attempts to position agentic capability as a standard expectation across all price tiers. According to TechCrunch, the model is designed to plan, utilize tools like browsers and terminals, and operate autonomously. Anthropic claims that Sonnet 5 can perform tasks at a level that, only months prior, required larger, more expensive models. This release arrives as other major AI labs pivot toward agentic workflows; OpenAI recently launched a preview of its GPT-5.6 Sol, and Google introduced Gemini 3.5 Flash in May.
The primary differentiator for Sonnet 5 is its cost structure relative to its performance. Starting Tuesday, the model is available across all subscription plans and serves as the default for free and Pro users. Pricing is set at $2 per million input tokens and $10 per million output tokens until August 31, after which costs increase to $3 and $15 per million tokens, respectively. TechCrunch reports that this pricing undercuts Opus 4.8, OpenAI’s GPT-5.5, and Google’s Gemini 3.1 Pro, though it remains more expensive than Google’s Gemini 3.5 Flash.
In terms of performance, Anthropic asserts that Sonnet 5 offers significant improvements over its predecessor, Sonnet 4.6, which debuted in February. On benchmarks for agentic coding, Sonnet 5 achieved a score of 63.2%, trailing Opus 4.8’s 69.2% but outperforming Sonnet 4.6’s 58.1%. While Opus 4.8 remains the preferred choice for complex judgment calls and deep research, Anthropic suggests that Sonnet 5 provides a higher-quality, lower-cost alternative for developers. Early feedback from users like Zapier senior engineer Daniel Shepard indicates the model can successfully complete multi-step tasks that previously stalled.
Safety remains a central focus of the release, with Anthropic reporting that Sonnet 5 demonstrates a lower rate of undesirable behaviors compared to Sonnet 4.6. The company states the model is more effective at refusing malicious requests, resisting prompt-injection attacks, and avoiding sycophancy. However, Anthropic acknowledges that Sonnet 5 does not match the safety profile of its more advanced models, such as Opus 4.8 or the Claude Mythos Preview, particularly regarding the ability to perform dangerous cybersecurity tasks. Fabian Hedin, co-founder of Lovable, noted that the model refuses unsafe requests consistently.
As the industry shifts toward agentic AI, the competition is increasingly defined by the balance between operational costs and reliability without human intervention. Anthropic frames the choice between Sonnet 5 and Opus 4.8 as a trade-off where users can adjust effort levels to suit their specific performance and budget requirements. For businesses, the utility of these models is being measured by their ability to manage complex, multi-step workflows—such as updating database records or executing API-based tasks—without the need for constant oversight.