The formula for a Claude feature's run rate is honest arithmetic: average input tokens per request × input price, plus average output tokens per request × output price, times monthly request volume. Everything hard about budgeting hides inside "average" and "volume." This worksheet walks a concrete example — a support-ticket summarizer moving from pilot to production — using only officially published list prices.
Step 1: Measure tokens per request from the pilot
Don't estimate token counts; measure them. Every API response includes a usage object with input_tokens, output_tokens, and cache fields — log them during the pilot and take the mean and the p95. For pre-launch estimates of prompt changes, the free token counting endpoint accepts your exact system prompt, tools, and messages. One trap when re-baselining: Claude Sonnet 5, Opus 4.7 and later, and Fable 5 use a newer tokenizer producing roughly 30% more tokens for the same text than earlier models, so counts measured on an old model don't transfer.
Suppose the pilot measures: 3,000 input tokens per request (system prompt + ticket text) and 400 output tokens (the summary).
Step 2: Multiply by list prices
On Claude Sonnet 5 at the standard rate of $3 per million input tokens and $15 per million output tokens, one request costs:
- Input: 3,000 × $3 / 1,000,000 = $0.009
- Output: 400 × $15 / 1,000,000 = $0.006
- Total: $0.015 per request
At a production volume of 200,000 requests per month, that's 600M input tokens ($1,800) plus 80M output tokens ($1,200) — a run rate of $3,000/month. Note the timing wrinkle: through August 31, 2026, Sonnet 5 has introductory pricing of $2/$10, which makes the same workload $2,000/month — budget at the standard rate so September doesn't look like a regression.
Step 3: Apply the modifiers you'll actually use
| Scenario (same workload) | Monthly cost |
|---|---|
| Sonnet 5, standard rates | $3,000 |
| + prompt caching: 2,500 of the 3,000 input tokens are a shared prefix served from cache at 0.1x ($0.30/MTok) → 500M cached tokens = $150, 100M uncached = $300, output $1,200 | ≈ $1,650 (plus a small cache-write cost) |
| Batch processing (async, 50% discount; most batches finish within an hour) | $1,500 |
| Claude Haiku 4.5 ($1/$5), if quality holds on evals | $1,000 |
Modifiers can stack — batch and caching discounts combine — and platform surcharges go the other way: regional endpoints on Bedrock and Google Cloud carry a 10% premium over global, and US-only inference (inference_geo: "us") applies a 1.1x multiplier on supported models. Marketplace list prices match first-party list prices across the 3P platforms, so this worksheet is platform-portable; only the discount negotiation differs.
Step 4: Forecast volume honestly
Token math is precise; volume is the guess. Write the volume assumption down as a range with a named driver ("40% of the 500K monthly tickets opt in → 200K requests"), and give finance three scenarios — pilot-observed adoption, plan-of-record, and a p95 ceiling. Then re-forecast monthly from actuals. Volume is also where growth hides: multi-turn features grow tokens per request as conversations lengthen, not just requests per month.
Step 5: Cap the downside
A forecast is a promise only if something enforces it. On the Claude API, set an organization spend limit below your tier cap and put the feature in its own workspace with a workspace-level spend limit — usage pauses rather than overruns. Rate limits (e.g., Start-tier Sonnet 5 at 2,000,000 input tokens per minute) bound the burn rate independent of budget. On Bedrock, Vertex, Foundry, and Claude Platform on AWS there are no Anthropic-side spend limits — use your cloud's budgeting tools and anomaly alerting. And wire token-count regression tests into CI so a prompt edit can't silently triple the per-request cost your whole forecast rests on.
Where to go next
Deepen the per-line-item understanding with anatomy of a Claude token bill, and connect spend to value with unit economics for LLM features.