Anthropic's Claude Sonnet 5 Targets Agent Builders on Price and Terminal Skill

Anthropic released Claude Sonnet 5 on June 30, positioning it as its most agentic mid-tier model and, per TechCrunch, a cheaper way to run agents in production.

The figure that matters most to engineers is Terminal-Bench, where Sonnet 5 scores 76.1 percent against Sonnet 4.6's 55.4 percent. That is a gain of more than 20 points on the benchmark that measures autonomous command-line work, according to MarkTechPost. On SWE-bench Pro it reports 63.2 percent, above Sonnet 4.6 at 58.1 percent but still below Opus 4.8 at 69.2 percent. On the OSWorld-Verified computer-use suite it posts 81.2 percent.

These are vendor-published figures on public benchmarks, not independent reproductions, and the more important difference is economic rather than a gain in peak capability. Sonnet 5 launches at $2 per million input tokens and $10 per million output tokens through August 31, roughly a third of the flagship Opus rates. The argument is that agent loops that were too expensive to run at Opus prices become viable at Sonnet prices, with only a modest capability gap.

The broader signal is where the training effort went. A 20-point gain concentrated in terminal and tool-use tasks, rather than general reasoning, shows that labs are now optimizing specifically for the long-horizon agentic workloads that coding assistants and autonomous systems depend on.