Anthropic releases cheaper, more agentic Claude Sonnet 5

Anthropic has released Claude Sonnet 5, a new midsize model designed to be more powerful and agentic than its predecessors. In this context, agentic refers to AI models capable of planning, using tools like browsers and terminals, and acting autonomously. According to Anthropic, the model can make plans, use tools, and run autonomously at a level that previously required larger and more expensive models. On Tuesday, Claude Sonnet 5 became the default model for free and Pro plans.

The model is positioned as a cost-effective alternative to larger models, including Anthropic’s own Opus 4.8, OpenAI’s GPT-5.5, and Google’s Gemini 3.1 Pro. Anthropic stated that while Opus 4.8 remains the preferred model for tasks requiring higher accuracy, Sonnet 5 offers developers a lower-priced option of higher quality than previously available. Users can adjust the effort level between the two models to find the right balance of cost and performance.

Claude Sonnet 5 pricing is structured as follows:

• Through August 31: $2 per million input tokens and $10 per million output tokens.
• After August 31: $3 per million input tokens and $15 per million output tokens.

Anthropic reports significant improvements in agentic performance for Sonnet 5 compared to Sonnet 4.6, which was released in February. On an agentic coding benchmark, Sonnet 5 scored 63.2%, compared to Sonnet 4.6’s 58.1% and Opus 4.8’s 69.2%. Testers have reported that the model successfully completes complex, multi-part tasks that previously stalled. Daniel Shepard, a senior engineer at Zapier, stated: “We handed Claude Sonnet 5 a two-part job — update Salesforce account tiers, send a launch announcement to enterprise contacts — and it finished end to end. That used to stall halfway. For day-to-day automation, it’s a no-brainer. ”

On safety, Anthropic reports that Sonnet 5 exhibits a lower rate of undesirable behaviors, such as cooperation with misuse and deception, compared to previous versions. The model is better at refusing malicious requests and sidestepping hijack attempts in prompt-injection attacks. It also hallucinates and engages in sycophantic behavior at a lower rate than Sonnet 4.6. However, it remains less capable than Opus 4.8 at performing dangerous cybersecurity tasks, and is not on the same level regarding misaligned behavior. Fabian Hedin, co-founder of Lovable, stated that Claude Sonnet 5 refuses unsafe requests cleanly and consistently, adding that a model capable of refusing unsafe requests is as important as one that knows how to build.

Why it matters

Claude Sonnet 5’s release confirms that agentic capability is now the baseline expectation for AI models across price tiers. The core market competition has shifted from which provider can deliver agentic work to how cheaply and reliably they can execute it without human oversight.