Anthropic launched Claude Sonnet 5 on June 30 as its most agentic Sonnet-class model. The company says it can plan, use browsers and terminals, write and debug code, and sustain longer tasks that previously pushed users toward more expensive Opus models. It is the default for Free and Pro plans and is also available to Max, Team, and Enterprise users, Claude Code, and the Claude API.

API pricing is part of the launch story. Anthropic set an introductory price of $2 per million input tokens and $10 per million output tokens through August 31, 2026. The stated standard price after that is $3 per million input tokens and $15 per million output tokens. Teams evaluating the model should calculate both launch-period and steady-state costs.

A wider cost-performance range

Anthropic presents Sonnet 5 as a substantial improvement over Sonnet 4.6 and, at higher effort settings, competitive with Opus 4.8 on some agentic search and computer-use tasks. That does not mean the smaller model replaces Opus everywhere. It means developers can trade more inference effort for capability across a wider range of price points.

  • Use the API model identifier claude-sonnet-5 and pin production behavior rather than relying on a moving default.
  • Run tests at the effort level you can afford; benchmark curves can change when the reasoning budget changes.
  • Include tool-call failures, browser mistakes, and incomplete task termination in the acceptance rate.
  • Model-generated code still needs tests, review, secrets controls, and a rollback path.

Benchmark claims need a local test

The launch comparisons are Anthropic’s evaluations. Some use external benchmarks, but harness choices, tool environments, retries, and effort settings influence results. Early-access testimonials also describe selected customer experiences rather than a representative audit. A team should reproduce the workflow that matters to it before changing routing or budgets.

The most revealing evaluation is completion quality over an entire task. Give Sonnet 5 a realistic repository, a clear test suite, constrained credentials, and a review rubric. Track whether it notices failures, recovers from tool errors, and stops at the right point. A model that produces a plausible patch quickly but needs extensive cleanup may be less economical than its token price suggests.

Sonnet 5 is important because strong agentic behavior is moving into a cheaper, broadly available tier. If the reliability holds outside Anthropic’s demonstrations, more organizations can run longer delegated workflows. The opportunity and the risk are the same: a model that can take more actions also needs better permission boundaries and supervision.