Anthropic Ships Claude Sonnet 5, Pushing Near-Flagship Agent Performance Down the Price Curve

Anthropic released Claude Sonnet 5 on June 30, presenting it as its most capable mid-tier model for autonomous tasks and pricing near-flagship capability well below its Opus line. The offer to engineers is direct. It provides planning, tool use, and long-horizon autonomy that a few months ago required larger models, at a lower token price.

The published numbers support a real advance for the Sonnet tier without matching the top model outright. On SWE-bench Verified, Sonnet 5 scores 72.7%, up from 62.3% for Sonnet 4.6 and short of Opus 4.8 at 79.4%. On Terminal-Bench 2.1, a test for command-line agents, Sonnet 5 reports 80.4%, ahead of Opus 4.8's 74.6%. On OSWorld-Verified computer use it trails the flagship, 81.2% to 83.4%.

Pricing is the more important factor. Anthropic set introductory rates of $2 per million input tokens and $10 per million output tokens through August 31, rising afterward to $3 and $15. That places frontier-adjacent agent performance at roughly a fifth of typical flagship output pricing, the figure that determines whether autonomous coding loops are economical at scale.

These figures come from Anthropic's own release and third-party compilations, not independent reproduction. The Terminal-Bench result reverses the usual order between Sonnet and Opus, which is notable and worth checking against outside runs. A mid-tier model beating the flagship on any agent benchmark is the kind of claim that deserves scrutiny.