New Claude models consume more tokens but cost less per solved task

A benchmark of 10 Terminal-Bench tasks run through Claude Code across Sonnet 4.6, Opus 4.7, and Opus 4.8 shows a shift in efficiency. Each run was instrumented using OpenTelemetry and SigNoz.

In terms of accuracy, Sonnet solved 5 tasks, Opus 4.7 solved 7, and Opus 4.8 solved 8. While Sonnet was the cheapest in raw cost at $6.30, Opus 4.8 proved cheaper than Opus 4.7 at $8.06 versus $8.51.

The critical metric is cost per solved task: $1.26 for Sonnet, $1.22 for Opus 4.7, and $1.01 for Opus 4.8. Opus 4.8 consumed the most tokens at 6.06 million compared to 2.88 million for Sonnet, but nearly all were cache reads, the least expensive category.

Tool error rates dropped from 1.72% for Sonnet to 0.85% for Opus 4.8. Sonnet failures were primarily timeouts, while the single Opus 4.8 failure was a fast, incorrect answer on a complex task.