What a Month of Claude Code Costs
The three variables that decide Claude Code spend — daily turns, re-sent context, cache hit rate — and a free calculator that models them. No signup.
Every team adopting Claude Code hits the same question from finance: "what will this cost per month per developer?" The honest answer — "somewhere between $20 and $2,000 depending on usage" — doesn't survive a budget meeting. The number isn't unknowable; it's just driven by variables most estimates never model. This post lays out the model, because the model matters more than any single number — and a free calculator at the end runs it for you.
Disclosure up front: I operate Teapot Router, an API gateway for teams running GPT and Claude Code. The calculator is free, needs no signup, and the cost model below applies to any Anthropic-compatible endpoint — nothing here depends on using ours.
Three numbers decide the answer
Agent bills aren't shaped like ordinary API bills. Before any price comparison, three variables settle most of the outcome, in order of impact:
- Turns per day. Agent workflows are many small requests, not few big ones — a heavy day of refactoring runs hundreds of turns, each one a billed request.
- Re-sent context. Every turn re-sends the system prompt, file context, and tool definitions. Your "input tokens" are mostly the same tokens, over and over — message 15 is billed against everything the agent re-read to answer it.
- Cache hit rate. The variable most teams never model. Cached reads bill at a fraction of list price and agent workloads are extremely cache-friendly — two teams with identical turn counts can differ several-fold in cost purely on cache behavior.
Why list-price comparisons mislead
That third variable is why provider comparisons at list price are nearly meaningless. An endpoint quoting full list price with honest cache pass-through can beat one quoting a discount that caches nothing. The only comparison that survives contact with a real invoice is the effective rate: what you actually paid per token after caching — which is exactly what a per-request ledger shows, and exactly what the calculator lets you project before you commit.
The calculator: model it, don't guess it
You set the variables above — daily turns, typical context size, output size, cache hit rate — and it projects a monthly cost per developer. More useful than the single number is the sensitivity view: the difference between a 60% and a 90% cache hit rate on your shape of workload is usually the largest lever you actually control. It's free with no signup and no email wall: https://teapotrouter.com/en/claude-code-cost-calculator — a cost tool that's a lead-gen trap is worse than no tool.
Where the model comes from
Enterprise customers kept asking for cost projections before onboarding; we kept answering with spreadsheets; the spreadsheet became the calculator. Metering every request — model, tokens in/out, cache hits, exact cost — is what a gateway does anyway, so the calculator just exposes the same math the invoice uses. If you run agent workloads, it's free regardless of whether you ever look at the gateway itself.
Skip the folklore. Model turns, re-sent context, and cache hits, and the monthly number stops being a guess — run your own shape through the calculator, and bring the sensitivity table to the budget meeting instead of a range.
FAQ
Is the calculator free?
Yes — no signup, no email wall: teapotrouter.com/en/claude-code-cost-calculator.
Does it only work for Teapot Router users?
No. The cost model is Claude Code's token economics with Anthropic list pricing; it's useful with any Anthropic-compatible endpoint.
What inputs does it need?
Daily turns per developer, typical context size, output size, and cache hit rate — with sensible defaults drawn from common agent workloads if you don't know yours yet.