What AI Features Cost to Run Per Month: Tokens, Models and Limits

An AI feature in an MVP costs between a few dollars and a few hundred dollars a month to run, and the number depends on three things: how many tokens each use sends and receives, which model tier you pick, and whether anyone set a cap. At 100 active users, document summaries, a chat assistant and semantic search together usually land under $100 a month on a mid-tier model and under $10 on a budget model. At 1,000 users the same three features run roughly $500 to $1,000 a month on a mid-tier model before caching, which can cut the chat part in half. The rest of this guide shows the arithmetic so you can repeat it for your own product, then lists the limits that stop a bill from surprising you. It covers only the AI line; for hosting, email and payments see what an MVP costs to run each month.

How AI billing works: tokens in, tokens out

AI providers charge per token, not per user or per request. A token is roughly three quarters of an English word, so 1,000 tokens is about 750 words. Every call has two meters. Input tokens are everything you send: the instructions your builder wrote (the "system prompt"), any document or search results you attach, and the user's message. Output tokens are what the model writes back. Output is priced four to five times higher than input, because generating text is more expensive than reading it.

Prices are quoted per million tokens. On a mid-tier model at $2 per million input tokens, sending a 3,000-word document costs about 0.8 cents; on a budget model at $0.10 per million it costs 0.04 cents. That twenty-times gap between tiers is the single biggest lever on your monthly bill.

Two more line items matter. Embeddings turn text into numbers so it can be searched by meaning, billed per million tokens at a fraction of a chat model's price. And prompt caching reuses an unchanged prefix (system prompt, a long document, earlier turns of a conversation) at roughly a tenth of the normal input price.

2026 price tiers you will actually choose between

Model names change every few months, so plan by tier rather than by name. As of October 2026 the major providers cluster into three bands, and the specific model can be swapped later without changing the design.

Two discounts apply across tiers. Cached input is usually billed at about 10 percent of the normal input price. Batch processing, where results come back within hours instead of seconds, is usually 50 percent off both meters. Neither needs a different vendor; both are turned on in code.

Worked example 1: document summaries

A user uploads a document and gets a short summary back. This is the simplest AI feature to cost because every use looks the same.

The tokens in one summary

• System prompt (how to summarize, tone, format): about 500 tokens.
• A typical 3,000-word document: about 4,000 tokens.
• The summary itself, 200 words plus formatting: about 300 output tokens.

That is 4,500 input tokens and 300 output tokens. On the mid tier the input costs $0.009 and the output $0.003, so one summary costs 1.2 cents. On the budget tier the same summary costs 0.06 cents. On the top tier it costs 6 cents, which is why the top tier is rarely the right default.

The monthly bill

Assume an active user runs 20 summaries a month. At 100 users that is 2,000 summaries: $24 on the mid tier, $1.20 on the budget tier. At 1,000 users it is 20,000 summaries: $240 on the mid tier, $12 on the budget tier. Summaries of short documents are a strong candidate for the budget tier, with the mid tier reserved for long or technical documents where accuracy matters more.

Watch the document length. A 30-page contract is 15,000 to 20,000 tokens, five times the example, and a user who uploads ten of those a day is not a $0.24 user any more. This is the first place a per-user limit earns its keep.

Worked example 2: a chat assistant

Chat is the feature most founders want and the one most likely to surprise them, because each new message resends the whole conversation. A ten-turn chat does not cost ten messages; it costs the first message once, the second twice, the third three times, and so on.

The tokens in one conversation

• System prompt with product knowledge and rules: about 1,500 tokens, sent on every turn.
• Average context per turn (prompt plus history so far): about 3,000 tokens.
• Average answer: about 250 output tokens.
• Turns per conversation: 10.

Over ten turns that adds up to about 30,000 input tokens and 2,500 output tokens. On the mid tier the input costs $0.06 and the output $0.025, so one conversation costs 8.5 cents. On the budget tier it costs 0.4 cents. With prompt caching on, 60 to 80 percent of that input is a repeated prefix billed at a tenth of the price, which brings the mid-tier conversation down to roughly 4 cents.

The monthly bill

Chat is where the mid tier is usually worth paying for, because users read every word and a budget model's flatter answers show. The right move is a tighter conversation, not a cheaper model: cap history at the last 8 to 10 turns, cap answers at 300 to 400 tokens, and cache the system prompt. Together these typically cut the chat line by 50 to 60 percent with no visible loss.

Worked example 3: semantic search over your content

Semantic search ("find the clause about late payment" across a library) has a one-time cost to index the content and a small cost per query. If you also generate an answer from the results, as most products do, add a chat-model call per search.

Indexing, paid once

A library of 10,000 pages at about 500 tokens a page is 5 million tokens. At embedding prices of $0.02 to $0.13 per million, indexing the whole library costs $0.10 to $0.65, and new content is embedded at the same rate as it arrives. The index lives in a vector database; managed services start free and run $25 to $70 a month past a few hundred thousand entries.

The tokens in one search with a generated answer

• Embedding the query: under 50 tokens, effectively free.
• Five retrieved passages of 500 tokens each: 2,500 input tokens.
• System prompt and the question: about 550 input tokens.
• The written answer: about 250 output tokens.

About 3,000 input and 250 output tokens per search. On the mid tier that is $0.006 plus $0.0025, so 0.85 cents a search. On the budget tier it is 0.04 cents. At 30 searches per user a month: 100 users cost $25.50 on the mid tier or $1.30 on the budget tier; 1,000 users cost $255 or $13. Search without a generated answer (just ranked results) costs almost nothing beyond the database.

The totals at 100 and 1,000 users

Put the three features together with the usage assumptions above (20 summaries, 8 conversations and 30 searches per active user each month), mid tier with caching on chat, and a small managed vector database.

Three things to take from the table. The bill scales with active users, not signups, so a product with 1,000 accounts and 150 active users pays the 100-user column. The gap between tiers is bigger than the gap between 100 and 1,000 users, so model choice per feature is the first decision, not the last. And even the expensive column is roughly 85 cents per active user per month, which only hurts if you charge nothing; should your MVP have a free trial works through that trade.

Six ways to cap spend before launch

None of these are optional for an AI MVP with a public signup. Ask to see where each one is set before the first real user.

1. A hard monthly budget at the provider. Every major AI console lets you set a spend limit that rejects requests once reached, plus email alerts at 50 and 80 percent. Set the limit at two to three times your expected bill. This is your backstop against bugs and abuse; it is not a plan.
2. A per-user allowance in your own code. For example 30 summaries, 10 conversations and 100 searches a day on the free plan, higher on paid. When a user hits it, show a message, not a silent failure. This is what turns a runaway user into a sales conversation.
3. Output caps on every call. Set a maximum answer length (300 to 500 tokens for chat, 400 for summaries). Output is the expensive meter and an uncapped model will happily write 2,000 tokens when 200 would do.
4. Trimmed history and cached prefixes. Keep the last 8 to 10 turns of a chat, summarize or drop the rest, and cache the system prompt. Together these cut chat input by more than half.
5. Route by task, not by habit. Use the budget tier for classification, tagging, routing and short summaries, and the mid tier only where a person reads the output directly. A simple rule in code decides per request.
6. Batch anything that can wait. Nightly digests, re-indexing, bulk tagging and report generation can run through the batch endpoint at half price. If a user is not waiting on the screen, it belongs in the batch.

Add one more that costs nothing: a daily spend number somewhere you will see it. A $20 jump on a Tuesday is a bug or an abuser, and you want to know by Wednesday.

What this means for your MVP scope and price

The monthly AI bill is small next to the build cost, but it continues forever and scales with success, so it belongs in your pricing model from day one. A product charging $19 a month with an AI cost of $1 to $3 per active user is healthy; a free product with the same cost is a slow leak. How to price your product before the MVP ships walks through setting a price that covers this line.

On the build side, the limits above are part of what a competent AI feature costs to build, not extras. What each MVP feature costs in 2026 shows where AI sits next to login, payments and admin, and fixed-price AI MVP development under $5,000 shows what a first AI MVP includes at that budget. For your own feature list, run it through the free MVP cost calculator.

Three things to confirm with whoever builds it: the AI provider account is in your name with your card; the model is set in one place so swapping tiers is a one-line change; and the per-user allowance and provider budget are configured and demonstrated before launch. At SquadPrime those are standard in any AI build: one fixed price from $4,000 agreed before we start, senior professionals doing the work, a private preview link you can click by day 5, live in 14 days or your invoice is zero, and all the code yours on delivery.

FAQ

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

For a typical MVP, between a few dollars and a few hundred dollars a month. At 100 active users, summaries, chat and search together usually cost under $100 a month on a mid-tier model and under $10 on a budget model. At 1,000 users, plan on roughly $500 to $1,000 a month on a mid-tier model before caching, or $50 to $100 on a budget model.

What is a token and why does it matter for the bill?

A token is the unit AI providers bill by. One token is roughly three quarters of an English word, so 1,000 tokens is about 750 words. Providers charge one price per million input tokens (what you send) and a higher price per million output tokens (what the model writes back), so long prompts and long answers are what drive the bill.

Can I set a hard limit on AI spend?

Yes. Every major provider lets you set a monthly budget in its billing console that stops requests once it is reached. On top of that, your app should enforce a per-user daily or monthly allowance, cap the length of each answer, and trim conversation history. Ask your builder to set all of these before launch, not after the first bill.

Should I use the cheapest model?

Not by default. The budget tier is fine for classification, extraction, routing and short summaries, and it is about twenty times cheaper than the mid tier. For customer-facing chat and anything that needs judgment, start on the mid tier and route only the simple requests to the budget model. A quality gap shows up in user trust faster than a cost gap shows up on your card.

Does prompt caching really cut the bill?

For chat and any feature that resends the same instructions or documents, yes. Cached input tokens cost about a tenth of the normal input price, and in a chat with a long system prompt and a growing history, 60 to 80 percent of the input is repeated. In practice this cuts the chat line by roughly half. It does nothing for output tokens.

Is the AI bill included in the fixed price of the MVP?

No. The fixed price buys the build. The monthly token bill is charged by the AI provider to an account you own, with your card on file, so it continues after handover and scales with your users. Budget for it separately, alongside hosting, email and payments.

All third-party names and trademarks belong to their respective owners and are used for identification only. Any market figures reflect our experience building MVPs and typical ranges as of October 23, 2026, and may change at any time. Spotted something outdated or inaccurate? Email talk@squadprime.com and we'll correct it promptly.

Related: What an MVP costs to run each month · What each MVP feature costs in 2026