Claude Sonnet 3.7

Running the realistic solo-developer day (1 medium feature + 5 small bug fixes + 4 PR reviews + 2 debug sessions + ~1500 lines of TypeScript + 1 large-doc read, 22 working days) on Claude Sonnet 3.7 costs about US$71/month. Heavier workloads scale proportionally; lighter workloads cost less.

200K tokens total, with up to 8K of output. That fits a few dozen source files in a single call.

Providers charge US$15.00 per 1M output tokens against US$3.00 per 1M input — output requires real compute, input comes mostly from cache. Coding agents read many files (input-heavy) and emit compact diffs (low output), so total spend is usually input-driven.

Cache reads typically cost only 10% of the regular input rate. On a coding-agent mix with 50% cache hits, that saves roughly 45% on input — which is about 38% off your total bill on input-heavy workloads. Anthropic models charge a one-time cache-write surcharge (25% over input) that pays for itself after 2–3 hits.

Extended-thinking / reasoning tokens are billed at the full output rate but never appear in your visible response. On hard agentic tasks they can double your output bill, lifting the monthly total by 20–30%. Enable thinking only when the standard response visibly fails.

Yes, if you can tolerate up to 24-hour turnaround: batch input/output are 50% cheaper than real-time rates. Perfect for nightly code reviews, bulk refactors or pre-merge analysis — wrong for inner-loop editing where you need an answer in seconds.