Calculator
What will this AI feature cost?
Estimate the monthly model bill before you build. Pick a feature like yours, adjust the numbers, and compare models side by side, with prompt caching and batch discounts included.
Like the “Catch me up” feature John shipped in Men of Hunger.
It can wait up to a dayBatch processing is half price.
Monthly estimate
200K requests · 1.3B tokens a month
Claude Haiku 5.5Fastest, cheapest
$160 /mo
$0.02 per user$0.80 per 1,000 requests
Claude Sonnet 5.5Everyday workhorse
$3,200 /mo
$0.32 per user$16.00 per 1,000 requests
Claude Opus 5.5Hard reasoning and agents
$6,400 /mo
$0.64 per user$32.00 per 1,000 requests
Claude Fable 5.1Most capable
$16,000 /mo
$1.60 per user$80.00 per 1,000 requests
Anthropic list prices, checked October 13, 2026. See the source. Assumes the cached prefix is rewritten once every ten requests (5-minute cache). Output tokens include any thinking the model does before it answers. Taxes, cloud-provider pricing and Priority Tier aren’t included.
How John keeps the bill down
- Start small, route up. Use the cheapest model that passes your quality checks, and send only the hard cases to a bigger one.
- Put the stable part of the prompt first. Instructions and reference docs that never change can be cached, and cached input costs a small fraction of the full price.
- Batch anything that can wait. Tagging, nightly digests and backfills rarely need an instant answer, and batch requests cost half.
- Watch output length. Output costs five times as much as input. Ask for exactly the format you need.
- Measure cost per finished task. A cheap request that needs three retries isn’t cheap. Track what one successful result costs.
Building an AI feature? John takes on select contract projects, from AI features to architecture reviews. See his work.