Anthropic released Claude Sonnet 5 on June 30, narrowing the performance gap between its mid-tier model and the more expensive Opus line.
The company describes Sonnet 5 as its most agentic Sonnet model, with stronger planning, tool use, browser and terminal work, coding and knowledge tasks. It is the default model for Claude’s Free and Pro plans and is also available in Claude Code, team products and the Anthropic API.
Anthropic set introductory API pricing at $2 per million input tokens and $10 per million output tokens through Aug. 31, after which the rates rise to $3 and $15. The company says higher effort settings can match Opus 4.8 on some tasks, though its own safety testing found Sonnet 5 substantially less capable than current Opus models in cybersecurity.
The middle tier gets more capable
Frontier releases receive attention because they define what is possible. Default models often matter more operationally because they determine what millions of users can afford to run every day. A less expensive model that can finish a multi-step task may create more business impact than a stronger model reserved for exceptional work.
Sonnet 5 also expands the number of cost-performance choices available to developers. Teams can vary effort based on the task rather than sending every request to the most capable tier. Routine research or code changes can stay on Sonnet; ambiguous or high-risk exceptions can move to Opus or a human reviewer.
That approach only works if routing is deliberate. A cheap agent that requires repeated retries, generates hidden rework or calls too many tools can cost more than a premium model that completes the job once. Buyers should measure task completion, latency, token use and review time together.
The cybersecurity result is another useful limit. Agentic performance is not one universal capability. A model can improve at browsing, coding and computer use without becoming equally strong at offensive security. Evaluations should match the actual workflow and the consequences of failure.
That is a healthier buying discipline than treating a model name as a single score.
For organizations already using Claude, the release makes a broader deployment model plausible: put Sonnet 5 into high-volume, bounded work and reserve the most expensive capability for the moments where it changes the outcome.
The headline is not simply that Sonnet became smarter. It is that more autonomous work is moving into the economic middle of the model market. That is where experiments become defaults.
About Kevin: Kevin Laird is an operator, systems strategist and Lean Six Sigma Black Belt focused on turning complex ideas into repeatable execution. His work spans AI implementation, automation, digital infrastructure, analytics and process design, with an emphasis on how organizations adopt new technology without losing clarity or control.