Usage-based pricing is the single biggest source of hesitation for people evaluating Claude Code, mostly because flat-fee SaaS pricing is what most people are used to.
Cost scales with how much the model reads and generates — larger codebases and longer sessions cost more than small, scoped tasks. This is why starting with a small, contained task isn't just good practice for learning safely, it's also the cheapest way to get a feel for real costs before scaling up.
Long-running sessions on large codebases, and repeatedly re-reading the same large files across many turns. Scoping a task tightly and working in a well-organized codebase both keep costs down.
Check Anthropic's current published pricing directly before starting, since usage tiers and included allowances change over time.
Costs track usage directly — a small, scoped task costs a small, predictable amount. The risk is mainly in leaving very long sessions running on very large codebases without checking in.