Inside Claude Sonnet 5: Anthropic’s new backbone for AI agents

Anthropic remains fully committed to the development of agentic AI. In this regard, its most powerful models belong to the Opus family, but the company is also focusing its efforts on the Sonnet line, designed for general use and centered on performance, speed, and cost-effectiveness. That lineup has just welcomed a new member: Claude Sonnet 5.

The developer introduced this AI as «the most agentic Sonnet model to date» and highlighted that it delivers performance similar to Opus 4.8, the most up-to-date version of Opus, but at a lower price point.

Claude Sonnet 5 has already been activated as the default model on the Free and Pro plans, and it is now available across all subscriptions with a special launch offer of $2 per million input tokens and $10 per million output tokens, valid through August 31. After that date, pricing will be set at $3 per million input tokens and $15 per million output tokens.

Introducing Claude Sonnet 5, our most agentic Sonnet yet.

It makes plans, uses tools like browsers and terminals, and runs autonomously at a level that just a few months ago required larger and more expensive models. pic.twitter.com/UKK8G7ww5h

— Claude (@claudeai) June 30, 2026

What Claude Sonnet 5 is like

Claude Sonnet 5’s agentic capabilities allow it to draw up plans, use tools such as browsers and terminals, and operate autonomously at a level that, until now, was only within reach of larger and more expensive models.

Anthropic shared Sonnet 5’s performance results across several benchmarks, along with a comparison against Sonnet 4.6 and Opus 4.8. As shown in the table below, Sonnet 5 represents a major upgrade over its predecessor, Sonnet 4.6, in reasoning, tool use, coding, knowledge, and agent performance, while its performance comes much closer to that of Opus 4.8, but at a lower cost.

Likewise, partners with early access say the AI can now detect and correct its own mistakes before considering a task complete, radically reducing friction in digital production environments.

For example, Fabian Hedin, co-founder of Lovable, said: «Claude Sonnet 5 gets more done with less. Same output quality, fewer steps to achieve it. On top of that, it rejects unsafe requests cleanly and consistently. At Lovable, we put powerful tools within reach of millions of developers. A model that knows when to say no is just as important as one that knows how to build».

Along the lines of what Hedin pointed out, it is worth noting that Anthropic reported that Claude Sonnet 5 «shows a lower overall rate of undesirable behaviors than Sonnet 4.6 and is generally safer for use in agent-based contexts». In addition, this AI is more efficient at rejecting malicious requests and resisting prompt injection attacks.

It is not a model designed for cybersecurity tasks

The company stated that «evaluations also show that it has a much lower capacity to carry out cybersecurity tasks than our current Opus models».

In other words, although the model can help you understand basic concepts or solve everyday, harmless cyber tasks, it falls well short when compared with the company’s “heavyweights” (Claude Opus 4.8 or Mythos 5). In real-world tests, such as attempting to hack the Firefox browser in a controlled environment, Sonnet 5 never managed to create a fully functional cyberattack.

While Anthropic mentions that Sonnet 5 is slightly better at this type of task than its previous version (Sonnet 4.6), achieving some “partial successes,” this is not because it was taught how to hack, but because the model now has stronger logical reasoning and greater general intelligence.

Likewise, to avoid complications or security issues, Anthropic has built into Sonnet 5 the same safety measures already enabled in Claude Opus 4.7 and 4.8, which can detect and block dangerous network use in real time.

In short, Anthropic is positioning Claude Sonnet 5 as an ideal and safe model for the mass consumer market. To achieve that, the company deliberately limited its offensive cybersecurity capabilities and, as an additional safeguard, natively integrated the same real-time protections that shield its high-end enterprise models.

Photo: generated with ChatGPT Images 2.0