Opus 5.5 Cut Coding Agent Costs by Half Despite Higher List Price Than Sonnet 5

A developer running multi-agent coding workflows on Claude models tracked five days of API spending between September 20–25, comparing Sonnet 5, Opus 5, and Opus 5.5 across 203 implementer pull requests and 411 reviewer runs. Although Opus 5.5 is listed at twice the per-token price of Sonnet 5, it cost roughly half as much in practice, delivering an estimated $0.39 per 100 changed lines versus $1.57–$2.13 for Sonnet 5. The key driver was not output token pricing but the number of API calls per task: Opus 5.5 required only 6 calls per 100 lines compared to 30–38 for Sonnet 5, and cached reads — which made up 68% of spending in one measured window — cost the same $0.20 per million tokens on both models. The developer also found that raising Sonnet 5's effort level increased cost by 35–51% per line with no clear improvement in first-review pass rates. Following the analysis, several workflow changes were adopted, including capping context windows at 400K tokens and extending reviewer cache duration, which together further reduced spending.