Anthropic's Claude Sonnet 5 posts 92.4% on SWE-bench Verified, narrowing the price-to-capability gap

Anthropic released Claude Sonnet 5 on June 30, positioning it as a lower-cost model for agents and coding that performs close to the company's flagship, Opus 4.8. On SWE-bench Verified, the standard patch-generation benchmark, third-party write-ups report Sonnet 5 at 92.4%, well above earlier Opus releases and above competing frontier models on the same harness.

The more interesting number is agentic. On Terminal-Bench 2.1, which measures multi-step command-line task completion, Sonnet 5 is reported at 80.4%, ahead of Opus 4.8's 74.6%, a result in which Anthropic's mid-tier model scores higher than its flagship on the same evaluation. On the harder SWE-bench Pro, Sonnet 5 sits at 63.2%, still below Opus 4.8 at 69.2%, so the ordering depends heavily on which benchmark you weight.

The pitch to engineers is economic rather than a raw capability jump. TechCrunch frames the launch as a cheaper way to run agents, and The New Stack notes the model is priced to close the gap with Opus and is discounted through August. For workloads that generate many tool calls, the cost of each token compounds, so a model at 90-plus percent of flagship quality at a lower rate changes the economics of running an agent continuously.

Two cautions. The headline scores now circulating come largely from Anthropic's own materials and vendor comparison posts, not from independent reproduction, and because scores on SWE-bench Verified now cluster near the low 90s, small differences are hard to interpret. What matters is the pattern. Anthropic is optimizing a cheaper tier specifically for code and agents, exactly where the competitive pressure is highest.