AI Cost Calculator
Your product isn't one cost mechanism — the same request carries memory, RAG, tools, images and hidden reasoning. This prices the whole stack on one stream (no double-counting) and ranks your biggest lever.
For engineers and founders pricing a whole AI product — models one request stream with every mechanism layered on (memory, RAG, tools, images, reasoning) instead of five separate calculators that double-count the base tokens.
Model prices from OpenRouter · updated 2026-07-13
01 Base conversation
$2/M in · $10/M out · cache reads $0.2/M
02 What rides on every requestset to 0 / 1 for anything you don't use
03 Naive estimate vs whole stack
Your stack costs 1.3× the “prompt + answer” estimate — RAG retrieved context is the biggest lever.
Without caching this stack would run $8,235/month — caching the stable prefix is doing heavy lifting.
Where the money goes
Deep-dive calculators for each line
Prices from OpenRouter, snapshot 2026-07-13, synced daily. One request = (system + tools + history) × [cache-read when caching, else input] + (user + RAG + images) × input + answer × (1 + reasoning ratio) × output, per 1M tokens; history = (user + answer) × (turns−1)/2. Rows sum to the total. Image and tool-schema token counts are your inputs — compute exact figures on their dedicated calculators. All math runs in your browser.
How the math works
A real AI product isn't one cost mechanism — the same request carries the base conversation plus whichever of these ride along: conversation memory, RAG context, tool schemas, image inputs, hidden reasoning. Adding up five standalone calculators would double-count the base tokens, so this page models ONE request stream and layers each mechanism on as its own line.
Same baseline, one identity: the naive "prompt + answer" estimate for your AI product is $0.005900/request ($2,950/month). Layer the real stack on — conversation memory replay $193, RAG retrieved context $1,200, the tool schema tax $216 — and the real bill is $0.007677/request, $3,839/month (1.3× the naive number). Every line is printed below and the rows sum exactly to the total.
Caching is modeled honestly by splitting the input: the stable prefix (system prompt, tool schemas, conversation history) re-reads at $0.2/M, while per-request fresh content (the user message, retrieved RAG chunks, images) always bills the full $2/M — retrieval results change every query, so they never cache. Without caching this stack would run $8,235/month instead of $3,839.
The biggest lever at these defaults is RAG retrieved context at $1,200/month — attack that line first (each mechanism has its own dedicated calculator for the deep dive). One scope note: agentic multi-step workflows re-send accumulated context on every step, a different traffic shape from the per-request stream modeled here — price those on the agent cost calculator instead.
Each line reuses the same tested engines as its dedicated calculator (conversation replay, tool schema tax, reasoning ratio, etc.); prices sync daily from OpenRouter (updated 2026-07-13). Token counts for images and tool schemas are your inputs — compute exact figures on the vision and tool-schema calculators, then carry them here. All math runs in your browser.
Frequently asked questions
What does your AI product really cost per month?
$3,839/month at these defaults — $0.007677/request across 500,000 requests. The naive "prompt + answer" estimate says $2,950; the difference is the stack riding on every request (conversation memory replay $193, RAG retrieved context $1,200, the tool schema tax $216). Tune the sliders to your real traffic.
Why not just add up the separate calculators?
Because they'd each re-count the base prompt and answer — the mechanisms share one request, they don't run as separate traffic. This page prices a single stream: base conversation once, then each mechanism as an additive line. The rows sum exactly to the total, so nothing is counted twice.
Which part of my stack should I optimize first?
At these defaults: RAG retrieved context ($1,200/month, the largest line after the base). Each line links to a dedicated calculator with the levers for that mechanism — caching, trimming, routing, or effort tuning depending on the line.
How does prompt caching interact with the stack?
Only the stable prefix caches: system prompt, tool schemas, and conversation history re-read at the discounted rate. The user message, RAG retrieved chunks, and images change every request, so they always pay full input price. Here caching cuts the stack from $8,235 to $3,839/month.
Where do I get the token counts for images and tool schemas?
They're provider-specific: compute exact image tokens on the vision token calculator (each provider tokenizes images differently) and tool schema tokens on the tool schema calculator (roughly 150–250 tokens per tool). Then carry the numbers here. RAG tokens ≈ chunks retrieved × tokens per chunk.
Are these prices current?
Model prices sync daily from OpenRouter (updated 2026-07-13). This page adds no separate price source — it composes the same tested per-mechanism engines used across the site, so it stays accurate as prices change automatically.