Most developers discover their Claude costs after the first invoice arrives.
Start with token counts. A 1,000-token prompt plus a 1,000-token response costs about $0.03 at current Claude rates. Multiply by daily volume: 100 prompts × 2,000 tokens = $3/day, or $90/month. This is the baseline for a lightweight coding assistant (case-006: a developer paid $145 before, $29 after).
For repetitive tasks like customer support replies, token reuse compounds. A small SaaS team (case-002) ran 50,000 auto-replies/month at 500 tokens each. Their raw cost was $1,840; with identical quality, it dropped to $287. The math: 50,000 × 500 tokens × $0.00003 = $750 — but real-world overhead (retries, context windows) pushes it higher. Assume 2–3× your token estimate for production systems.
Agentic workflows are the worst offenders. An agency (case-004) built a research-draft-critique loop that generated 8,300 tokens per cycle. At 300 cycles/month, their bill was $2,490. The same output cost $498 after. The lesson: every agent step multiplies tokens. Count each iteration, not just the final output.
Audit your own Claude usage. Paste your last 30 days at aiusage.ai — no signup for the number.