Depends entirely on whether you pay per-token through the API or through a flat Pro/Max subscription — and on how much of your context gets cached. Here's the real math for each.
| Usage level | Est. tokens/day | API cost/day (Sonnet 5) | API cost/month |
|---|---|---|---|
| Light (a few small tasks) | ~150K | $0.45–$0.90 | ~$15 |
| Normal (daily driver, some caching) | ~800K | $2–$4 | ~$70 |
| Heavy (long agentic sessions, multiple repos) | ~3M+ | $8–$15 | $200+ |
Ranges assume a realistic mix of input/output tokens and partial prompt-caching (session context reused across turns). Actual cost swings a lot with codebase size, how much the agent re-reads files, and cache hit rate — see the caching math for why the same session can cost 5-10x more with caching off.
The subscription plans (Claude Pro, Claude Max) bundle Claude Code usage into a flat monthly fee with soft usage limits, which is predictable but can throttle a genuinely heavy day. The API is metered exactly to the token and has no cap, which is cheaper for light/normal use and more expensive — but uncapped — for sustained heavy use. Rule of thumb: if you use Claude Code most of the workday, every day, a flat subscription is usually the better deal; if usage is bursty or occasional, pay-as-you-go API billing is usually cheaper.
Three things move the number far more than which model you pick:
| Factor | Effect |
|---|---|
| Context re-reading | An agent that re-reads the same large files every turn pays full input price each time — caching (or reading less) cuts this dramatically. |
| Session length | Longer sessions accumulate more context in the transcript, and every turn re-sends that history unless it's cached. |
| Model choice | Sonnet 5 is the default and priced in the middle; switching to Haiku 4.5 for simple subtasks (not the whole session) is the highest-leverage cost lever after caching. |
The rough tiers above are a starting point, not your bill. The full calculator lets you plug in your actual input/output token counts and caching setup for an exact per-session and per-month number, and the model pricing comparison shows what switching models alone would save. If you're mostly using Claude Code interactively rather than building against the raw API, also check whether a flat-rate Pro/Max plan beats metered API cost for your usage level.