Ask AI — Cost Calculator

Per-Request Token Breakdown

Every time someone asks a question, these tokens are sent to and received from the AI:

Component	Tokens	Direction	Notes
System prompt	~500	Input	Instructions for the AI historian role
Database context	~2,000	Input	Matched family records (capped at 15 people)
Site help (sometimes)	~400	Input	Only included for "how do I" questions
User's question	~50-100	Input	The actual question text
Conversation history	~0-1,500	Input	Up to 6 prior exchanges for follow-up questions
Total Input (typical)	~2,600	Input	First question, no history
Total Input (follow-up)	~4,000	Input	With conversation history
AI response	~400	Output	2-4 paragraph answer

Model Comparison — Cost per Request

Haiku 4.5

Recommended

Input: $1.00 / MTok

Output: $5.00 / MTok

Cost per question:

$0.0046

~217 questions per dollar

Sonnet 4

Output: $15.00 / MTok

Cost per question:

$0.0138

~72 questions per dollar

Opus 4

Premium

Input: $5.00 / MTok

Output: $25.00 / MTok

Cost per question:

$0.023

~43 questions per dollar

Why Haiku? For a genealogy Q&A where we pre-gather the relevant family records from the database, Haiku 4.5 gives answers that are just as good as Sonnet — but it's 3.7x cheaper and responds noticeably faster. We switched from Sonnet to Haiku in this update.

Scaling Calculator

Weekly Active Users

100 users

Avg Questions per User per Week

3 questions

Model

Cost at Scale — All Models (3 questions/user/week)

Weekly Users	Q/Week	Haiku 4.5 /mo	Sonnet 4 /mo	Opus 4 /mo

API Credit → Real Dollars

Anthropic bills in dollars directly (no credit system). Here's how far each dollar goes:

Monthly Budget ($)

$25 / month

Model

Before vs After This Update

Metric	Before (Sonnet 4)	After (Haiku 4.5)	Improvement
Cost per question	$0.0138	$0.0046	3x cheaper
100 users/week (monthly)	$16.56	$5.52	Saves $11.04/mo
1,000 users/week (monthly)	$165.60	$55.20	Saves $110.40/mo
First-token latency	~3-5 seconds	~0.5-1 second	3-5x faster
Context payload	Uncapped (50-100KB)	Capped at 8KB	6-12x smaller
DB queries per request	Unlimited per-term	Capped at 15 matches total	Much fewer queries