trelvane
100% Free Forever 100% Private · Inputs Never Saved No Sign-Up Required

AI SaaS & RAG Production Unit Economics & API Cost Calculator

Model your startup’s monthly LLM COGS by Daily Active Users (DAU), requests per user, retrieved RAG chunk tokens, and output completion length across 11 frontier models.

Live BPE: 125 tokens · 47 words · 354 chars (2.83 chars/tok)
Override Input Tokens:
Quick Sample Scenario:
Calculating...
Monthly Production API Bill (900,000 Calls/Mo)
$850.50/ mo$0.00094 / call · $0.95 / 1K calls
Tokens/Call: 3,500 in + 700 outContext Used: 3.28% of 128K
All 11 Frontier Models Side-by-Side (900,000 Calls/Mo)
ModelPer 1K CallsMonthly Bill
Claude 4 Sonnet($3/$15)$21.00$18,900.00
GPT-5 / 4.1($2/$8)$12.60$11,340.00
GPT-4o($2.5/$10)$15.75$14,175.00
GPT-4o mini($0.15/$0.6)$0.95$850.50
OpenAI o3-mini($1.1/$4.4)$6.93$6,237.00
Claude 4 Opus($15/$75)$105.00$94,500.00
Claude 3.5 Haiku($0.8/$4)$5.60$5,040.00
Gemini 2.5 Pro($1.25/$10)$11.38$10,237.50
Gemini 2.5 Flash($0.15/$0.6)$0.95$850.50
DeepSeek R1($0.27/$1.1)$1.72$1,543.50
Llama 4 70B($0.59/$0.79)$2.62$2,356.20
Caching 75% of repeated system prompts saves $177.19/mo!

11-Model API Cost Comparison (3,500 Input + 700 Output Tokens)

Model Cost / 1K Calls Monthly Bill (900,000 Calls)
Anthropic Claude 4 Sonnet / 3.7 Sonnet $21.00 $18,900.00 / mo
OpenAI GPT-5 / GPT-4.1 Flagship $12.60 $11,340.00 / mo
OpenAI GPT-4o Multimodal $15.75 $14,175.00 / mo
OpenAI GPT-4o mini (High-Speed) $0.95 $850.50 / mo
OpenAI o3-mini / o4-mini (Reasoning) $6.93 $6,237.00 / mo
Anthropic Claude 4 Opus $105.00 $94,500.00 / mo
Anthropic Claude 3.5 Haiku $5.60 $5,040.00 / mo
Google Gemini 2.5 Pro (2M Context) $11.38 $10,237.50 / mo
Google Gemini 2.5 Flash (1M Context) $0.95 $850.50 / mo
DeepSeek R1 / V3 Reasoning API $1.72 $1,543.50 / mo
Meta Llama 4 70B / 405B (Groq / Together) $2.62 $2,356.20 / mo

Frequently Asked Questions

How do I estimate input tokens for a Retrieval-Augmented Generation (RAG) pipeline? ▾

A typical RAG query combines a 400-token system prompt + 5 retrieved vector chunks of ~500 tokens each (2,500 tokens) + a 100-token user question = ~3,000 input tokens per call.