How much does it cost to run an AI feature per user per month?

It depends on the model and how heavily each user uses it. As a worked example: 800 input plus 400 output tokens per request, 50 requests per user a month, on Claude Haiku 4.5 costs about $0.14 per user per month. Swap in your own token sizes, request rate, and model above for a figure tailored to your app.

What should I charge for my AI app to stay profitable?

Charge comfortably above your cost per user. The break-even price shown here is the raw LLM cost per user — the point where revenue equals the model bill. Real break-even is a little higher once you add payment-processor fees, hosting, and support, so a common rule is to price so the LLM bill is at most 20–30% of the subscription.

What is a good gross margin for an AI SaaS product?

Traditional SaaS runs 70–85% gross margin. AI features cut into that because inference is a real variable cost. If the LLM bill is 10–30% of your price you are in healthy territory (70–90% margin). If it climbs above 50% the model choice or usage pattern needs work — try a cheaper model or prompt caching before raising the price.

How do I calculate the break-even price for an AI subscription?

Multiply your cost per request by requests per user per month to get cost per user. That figure is the break-even price: below it every user loses money on the model bill alone. Cost per request is input tokens ÷ 1,000,000 × input price plus output tokens ÷ 1,000,000 × output price, using the vendor's per-million-token rates.

Which model gives the lowest cost per user?

For most chat and support workloads the cheapest seeded models are Gemini 2.0 Flash, GPT-5 nano, and GPT-4o mini, with Claude Haiku 4.5 close behind. It depends on your input/output token mix because output is priced higher than input. The comparison table above prices your exact workload across every model and sorts cheapest first.

Does prompt caching change my cost per user?

Yes. If a large part of each request is a repeated prompt or shared context, caching bills that portion at roughly 10% of the input price. Set the cached input share above and the cost per user drops accordingly. This tool applies a flat 10% cache-read rate and ignores the one-time cache-write surcharge, so treat cached figures as an optimistic estimate.

Why show the cost in Sri Lankan rupees?

Most founders here budget and price in rupees even though the API is billed in US dollars. Switch the currency toggle to rupees and the cost per user, break-even, and totals convert using an editable CBSL indicative exchange rate, while the dollar figure stays available alongside.

What costs are not included here?

This tool models the LLM inference cost per user only. It excludes hosting, vector-database storage, embeddings, payment-processor fees, fine-tuning, and image or audio tokens. Those have their own calculators on the site. Break-even and margin therefore reflect the model bill alone — your true break-even sits a little higher.

AI · SaaS unit economics

AI Cost Per User Calculator

Turn raw model token prices into the numbers that decide the business: your LLM cost per active user, the break-even subscription price, and the gross margin at any price you set — across Claude, GPT, and Gemini, in dollars and rupees. No signup, sources cited.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jul 8, 2026

Cost per user & margin

Model

Grouped by provider. Price shown is input / output per 1M tokens.

Currency

The dollar figures come first; rupees use your editable rate.

Avg input tokens / request

Prompt + context sent per request

Avg output tokens / request

Tokens the model generates

Requests / user / month

How often one user hits the model

Active users (MAU)

Monthly active users

Cached input share (%)

Repeated prompt served from cache

Your price / user / month ($)

Subscription price you charge

Quick presets

Cost per user / month

$0.14

$0.0028 per request

Break-even price / user

$0.14

Where revenue = LLM cost

Gross margin at $5.00

97.2%

$4.86 profit per user

Total monthly cost

$140.00

$1,680.00 / year

Where your price goes

LLM cost: $0.14 / userMargin left: 97.2%

Cost per user, every model

Model	Cost / user	Margin
GPT-5 nanocheapest OpenAI	$0.01	99.8%
Gemini 2.0 Flash Google	$0.012	99.76%
GPT-4o mini OpenAI	$0.018	99.64%
Llama 4 Scout (via Together AI) Meta (Together AI)	$0.019	99.62%
Mistral Small Mistral	$0.02	99.6%
Llama 4 Maverick (via Together AI) Meta (Together AI)	$0.0278	99.44%
DeepSeek V3 DeepSeek	$0.0328	99.34%
GPT-5 mini OpenAI	$0.05	99%
Gemini 2.5 Flash Google	$0.062	98.76%
DeepSeek R1 DeepSeek	$0.0658	98.68%
o3-mini OpenAI	$0.132	97.36%
Claude Haiku 4.5selected Anthropic	$0.14	97.2%
Mistral Large 2 Mistral	$0.20	96%
GPT-5 OpenAI	$0.25	95%
Gemini 2.5 Pro Google	$0.25	95%
GPT-4o OpenAI	$0.30	94%
Claude Sonnet 4.6 Anthropic	$0.42	91.6%
Claude Opus 4.8 Anthropic	$0.70	86%
o1 OpenAI	$1.80	64%

Total cost as you scale

100 users

$14.00

per month

1,000 users

$140.00

per month

10,000 users

$1,400.00

per month

100,000 users

$14,000.00

per month

Same per-user cost ($0.14) on the Claude Haiku 4.5 model, multiplied by each user count. LLM cost scales linearly with users — the margin question does not change as you grow.

Sources cited: Anthropic, OpenAI, and Google Gemini pricing pages (token prices, dated 2026-06-30) and the CBSL indicative USD→LKR rate — full links in the Sources & references section below. Break-even excludes payment-processor fees, so your real break-even is slightly higher. Cost-per-user is cross-checked against an independent per-1,000-token formula and both agree. Nothing leaves your browser.

How it works

A monthly API-bill calculator answers “what will I spend?” This one answers the question a founder actually loses sleep over: at my price, do I make money on each user, and which model keeps me in the black? It works from the same vendor token prices — input and output are priced separately per 1,000,000 tokens, with output almost always more expensive because the model spends compute generating each token.

Let inP and outPbe the model's input and output price per 1M tokens. The calculation runs in five steps:

Effective input price after caching. A cached share of the input is billed at about 10% of the input price, so effInP = inP × (1 − cachedShare × 0.90), with cachedShare between 0 and 1.
Cost per request. costReq = (inputTokens / 1e6) × effInP + (outputTokens / 1e6) × outP.
Cost per user per month. costUser = costReq × requestsPerUser. This is the headline figure. Total monthly cost is costUser × MAU, and the annual figure is twelve times that.
Break-even price. breakEven = costUser — the subscription price where revenue equals the LLM cost. Payment-processor fees push the real break-even slightly higher; that is out of scope for v1 and noted in the tool.
Gross margin. At your price P, margin = (P − costUser) / P. A negative result means the model bill exceeds your price — a per-user loss, rendered in red. Rupee figures multiply the dollar cost by an editable CBSL indicative exchange rate.

Every figure is cross-checked: the per-1,000,000-token math is re-derived through an independent per-1,000-token path and the two must agree to the cent, the same credibility check the site's tax calculator runs against the IRD formula. Token prices are the shared vendor pricing table (Claude, GPT, Gemini), last verified 2026-06-30; the formulas were reconciled 2026-07-08.

Worked examples

Each example works the math by hand. All three are encoded in the tool's self-test and reconcile to the figures shown, at Rs 300 per US dollar.

A — WhatsApp support bot on Claude Haiku 4.5 ($1 in / $5 out)

Workload: 800 in + 400 out tokens, 50 requests/user, 2,000 users, price $5, no caching
costReq = 800/1e6 × $1 + 400/1e6 × $5 = $0.0008 + $0.0020 = $0.0028
costUser = $0.0028 × 50 = $0.14/user (≈ Rs 42)
Monthly = $0.14 × 2,000 = $280.00 (≈ Rs 84,000); annual = $3,360
Break-even = $0.14/user; margin at $5 = (5 − 0.14)/5 = 97.2%

B — RAG assistant on GPT-5 ($1.25 in / $10 out)

Workload: 4,000 in + 800 out tokens, 200 requests/user, 500 users, price $9, no caching
costReq = 4,000/1e6 × $1.25 + 800/1e6 × $10 = $0.0050 + $0.0080 = $0.0130
costUser = $0.0130 × 200 = $2.60/user
Monthly = $2.60 × 500 = $1,300 (≈ Rs 390,000)
Break-even = $2.60/user; margin at $9 = (9 − 2.60)/9 = 71.1%

C — Same RAG assistant, but 60% of the input is cached

Caching lowers the input price: effInP = $1.25 × (1 − 0.6 × 0.9) = $1.25 × 0.46 = $0.575
costReq = 4,000/1e6 × $0.575 + 800/1e6 × $10 = $0.0023 + $0.0080 = $0.0103
costUser = $0.0103 × 200 = $2.06/user (down from $2.60)
Margin at $9 = (9 − 2.06)/9 = 77.1% — caching lifts margin about 6 points

Frequently asked questions

Sources & references

Token prices are transcribed from the vendor pricing pages above and last verified on 2026-06-30; the formulas and worked examples were reconciled on 2026-07-08. Prices move often, so re-check the live pages before committing a budget. The tool runs entirely in your browser — no inputs leave your device.

Related tools

LiveAI

AI Chatbot Cost Calculator

Estimate the monthly API cost of a multi-turn AI chatbot across Claude, GPT, and Gemini. Models the quadratic context re-sending that single-call calculators miss, with and without prompt caching, in USD and LKR.

Open tool

LiveAI

RAG Cost Calculator

Estimate the full one-time and monthly cost of a Retrieval-Augmented Generation pipeline — embedding, vector storage, retrieval, and LLM generation — across OpenAI, Anthropic, Google, Cohere and Pinecone, in USD and LKR. Every price sourced.

Open tool

LiveAI

AI Agent Cost Calculator

Estimate the real per-run, daily, and monthly cost of a multi-step LLM agent across Claude, GPT, and Gemini. Models the context accumulation single-call calculators miss — each tool result is re-sent every step — with caching, in USD and LKR.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, a pricing change, or want another provider added?

Email me at [email protected] — most fixes ship within 24 hours.