What an AI feature costs, request by request.
Set the traffic, the tokens and the model. You get the monthly bill, then what prompt caching, a small model for simple requests, batch processing and shorter answers take off it.
01Traffic and tokens
600,000 model calls a month02Model
Prices per million tokens, USDClaude API list prices (Large: Claude Opus 5.5, Medium: Claude Sonnet 5.5, Small: Claude Haiku 5.5). Prices checked on 9 November 2026. Set them to your provider's price list. Writing to the cache costs 125% of the input price; reading it back 5% on Opus 5.5 and Sonnet 5.5, 10% on Haiku 5.5. Batch is half price.
03Levers
Level 5 · architectureHow this is calculated
- Model calls a month: requests a day × 30 days × model calls per request.
- A call costs its fixed prompt and its variable input at the input price, plus its output at the output price, per million tokens.
- Prompt caching: the hit rate you set is the share of calls that read the fixed prompt from the cache, at the cache read price; the other calls write it again at the 5-minute cache write price, 125% of the input price.
- A small model for simple requests: that share of the calls is priced at the small tier's prices.
- Batch processing: that share of the calls costs 50% less, input and output. The discount stacks with caching.
- Shorter answers: the output tokens are cut by the percentage you set.
- Claude Haiku 5.5 is priced by prompt length: when the fixed prompt and the variable input add up to more than 100,000 tokens, its prices are 5 times higher.
- Each lever's saving is measured after the ticked levers above it, in the order of the list, so the savings add up to the total.
- Euros at a fixed 1 USD = €0.89, excluding tax. Not included: the minimum prompt length a cache needs, the 1-hour cache, US-only inference, the tokens tool use adds, fast mode and negotiated discounts.
The hit rate, the share of simple requests, the share of work that can wait and the output cut are your assumptions: the examples start from typical values, set them to your own traffic.
| Tier | Input | Output | Cache write, 5 min | Cache read | Batch input | Batch output |
|---|---|---|---|---|---|---|
| Large · Claude Opus 5.5 | $4 | $20 | $5 | $0.20 | $2 | $10 |
| Medium · Claude Sonnet 5.5 | $2 | $10 | $2.50 | $0.10 | $1 | $5 |
| Small · Claude Haiku 5.5 prompts up to 100,000 tokens | $0.10 | $0.50 | $0.125 | $0.01 | $0.05 | $0.25 |
| Small · Claude Haiku 5.5 prompts over 100,000 tokens | $0.50 | $2.50 | $0.625 | $0.05 | $0.25 | $1.25 |
Source: Anthropic, Pricing, Claude Platform docs ↗ (opens in a new tab). Example: the Support chatbot costs €5,927 a month as is and €1,523 with its levers.
The audit puts your own figures on it.
Free, 48 hours, from your billing exports and with your negotiated discounts. I'm paid 20% of the savings measured on your bill, for 12 months, and nothing if the bill does not go down.