Claude’s Sonnet 5 is built to do more on its own and cost you ...

What Happened

Anthropic has rolled out Sonnet 5, the latest iteration of its mid-tier Claude model, positioning it as the new default across the Free and Pro tiers while also extending availability to Max, Team, Enterprise, Claude Code, and the API. The company says the model is designed to operate with greater autonomy, taking on more multi-step tasks with less hand-holding than previous Sonnet versions required. Alongside these capability upgrades, Anthropic reports measurable safety improvements: Sonnet 5 hallucinates less and displays reduced sycophantic tendencies compared to its predecessor. Notably, the company also disclosed that the model is deliberately weaker at dangerous cybersecurity tasks than its flagship Opus-class models — a tradeoff Anthropic frames as intentional rather than a limitation of the technology.

Why This Matters

The framing here is significant for a few reasons. First, the push toward greater model autonomy reflects a broader industry trend: AI labs are racing to build systems capable of executing complex, multi-step workflows independently, whether that's coding, research, or agentic task completion. Sonnet's positioning as a workhorse model — available broadly rather than gated behind premium tiers — suggests Anthropic wants this autonomy-focused approach in the hands of as many users as possible, not just enterprise customers paying top dollar for Opus.

Second, the safety disclosures matter because they address two of the most persistent criticisms leveled at large language models: hallucination and sycophancy. Sycophancy — the tendency of models to tell users what they want to hear rather than what's accurate — has drawn increasing scrutiny as AI assistants are used for decision-making in professional and personal contexts. Anthropic quantifying improvement here, rather than simply asserting it, is a step toward more accountable reporting, though outside verification will matter.

The Cybersecurity Tradeoff

Perhaps the most interesting detail is Anthropic's admission that Sonnet 5 is deliberately less capable at dangerous cybersecurity tasks than Opus-class models. This is a notable acknowledgment: as models become more autonomous and widely deployed, the risk calculus changes. A model sitting behind a free tier, accessible to millions, poses a different threat profile than a premium model used by vetted enterprise clients. By intentionally limiting certain capabilities in the more widely distributed model, Anthropic appears to be building tiered risk management directly into its model lineup — reserving the most powerful (and potentially most dangerous) capabilities for contexts with presumably more oversight.

Broader Context

This release lands amid intensifying competition among AI labs to balance capability with safety credentials, particularly as regulators and enterprise customers demand more transparency. Whether tiered capability-safety tradeoffs become a standard industry practice, or whether they simply reflect Anthropic's particular safety-first branding, will become clearer as competitors respond.