Token Cost Calculator + Savings Tips

How Claude Code spend actually works

Claude Code bills you per token in and per token out, with a discount on tokens served from the prompt cache. That is the whole pricing model. Everything else - per-day cost, per-week cost, why your bill jumped last month - is just downstream of those three numbers. Once you internalize that the cost of a Claude Code session is roughly 'input tokens you pay full price for, plus input tokens you pay cache rate for, plus output tokens', the bill stops feeling mysterious and starts feeling like a knob you can turn.

The calculator above asks for six honest numbers - how long you work, how often you prompt, how much context you send, how much you get back, and how much of that context is repeated. From those six numbers it estimates your per-prompt, per-day, per-week, and per-month cost on three different models. The 'same workload, other models' panel tells you what you would pay if you switched - which is the single biggest cost lever most users miss.

The four levers that actually move the bill

• Model choice - Haiku is roughly 1/20th the cost of Opus per token. For most daily work, Sonnet is the right default; reach for Opus only when reasoning is the bottleneck.
• Cache hit rate - Anthropic caches your repeated context (CLAUDE.md, file contents, conversation history) at roughly 10% of normal input cost. Healthy sessions hit 30-50%. Bad sessions thrash and hit 0%.
• Input token budget per prompt - the single biggest knob you control. Most expensive prompts are expensive because they paste in a giant file that did not need to be there.
• Output token budget per prompt - the second biggest knob. Long, verbose answers cost five times what input does on every model. If you only need a diff, ask for a diff, not an essay.

Savings tips that actually work

1. Default to Sonnet, escalate to Opus on hard reasoning - do not default to Opus for everything 'just in case'. Most coding work fits in Sonnet's headroom.
2. Write a tight CLAUDE.md (under 200 lines). It is loaded into every session, so every line is a forever cost. The CLAUDE.md generator on this site enforces this by default.
3. Use the prompt cache deliberately. Reference the same files in the same order across prompts. Re-ordering breaks the cache; consistency lets it land.
4. Cap your output. End prompts with 'reply with only the diff' or 'reply in under 100 words' whenever the long-form answer is not the goal.
5. Plan first, then execute. A 200-token plan that catches a wrong direction is cheaper than a 5,000-token implementation in the wrong direction.
6. Trim attached file paths. Claude Code can attach huge files when asked to read them - half of cost overruns are 'I asked it to read this 80kb config and it included the whole thing twice'.
7. Compact sessions when they get long. A bloated session reloads the entire history each turn, which costs more than starting a fresh session would.
8. Watch the token meter in the CLI. Most users never look at it; the ones who do quickly internalize what their average prompt costs.

When raising the bill is the right call

Cost optimization should never become its own goal. Claude Code is cheap relative to the value it produces - if you are spending $100/mo and shipping work that would have taken you a week to do by hand, the bill is doing its job. The reason to track cost is not to minimize it. It is to notice the moments when spend goes up but output does not - that is the signal that something in your workflow drifted. Maybe a CLAUDE.md got bloated. Maybe you defaulted to Opus on a Tuesday and forgot. Maybe a session is reloading too much history. Fix the leak, not the budget.

What the calculator does not model

The estimator above is a first-order approximation. It assumes consistent prompt size, average cache behavior, and current pricing. It does not model: tool calls (which can spike a prompt's effective output), MCP-induced context (each MCP loads tokens), or long-context overage. Treat the monthly number as a planning estimate, not a guarantee. Track your actual usage for two weeks before locking in a budget.

What to do with the number

If the number is bigger than expected, the savings tips above land roughly 20-50% reductions in real workflows. If the number is smaller than expected, you might be under-using Claude Code - escalate to bigger tasks, run longer sessions, do more in the model and less in the editor. The Claude Code Club curriculum has a full module on workflow cost-efficiency for people building serious things with Claude Code. The calculator is the diagnostic; the curriculum is the prescription.