Claude Sonnet 5 arrives, and the mid-tier model now outscores Anthropic's own top model on agentic coding

Anthropic released Claude Sonnet 5 on June 30, positioning it as its most agentic mid-tier model and the default for free and paid users on claude.ai. The notable result is not the overall coding score but the internal ranking. On Terminal-Bench 2.1, Sonnet 5 scored 80.4 percent, ahead of the larger Opus 4.8 at 74.6 percent and Sonnet 4.6 at 67.0 percent. In this release, a Sonnet model scored higher than its larger Opus counterpart on a major agentic-terminal evaluation.

On the harder software-engineering benchmark, the ranking is as expected. Sonnet 5 reported 63.2 percent on SWE-bench Pro against Opus 4.8's 69.2 percent, roughly five points above the previous Sonnet. Anthropic also cited 81.2 percent on OSWorld-Verified for computer use and 84.7 percent on BrowseComp for agentic search. These are vendor-published figures on public benchmarks, and independent reproduction across held-out task suites has not yet been published, so treat the exact differences as claims rather than settled facts.

The commercial implication is clearer than the benchmark one. Anthropic set introductory pricing at 2 dollars per million input tokens and 10 dollars per million output tokens through August 31, rising to 3 dollars and 15 dollars afterward, far below Opus. For teams running agents at scale, a model that matches or beats the top model on terminal work at Sonnet prices changes the default routing decision, and it arrives as Anthropic is reported to be moving toward a public offering.