How Much Will Your AI Feature Cost at 1,000 Active Users?

That number will be wrong by a multiple, not by a few percent.
The prototype sends short prompts with no history. It makes one model call per question and never retries. Nothing runs in the background, and nobody runs evals against it. Production has all of that, and every piece of it bills tokens. None of it appears on the beta invoice because the beta didn't have it yet.
In the post on taking an AI prototype to production we priced one request and one active user. Here we take the same model to 1,000 active users and add what the prototype left out. Every number below is illustrative unless it links to a public price list. Swap in your own logs.
What we assume
The prices are A1's: $3 per million input tokens and $15 per million output tokens. That's the list price of Claude Sonnet 4.6 on Anthropic's pricing page as of September 2026. Other models and providers price differently. The method carries over.
An active user sends 20 requests a day, 30 days a month. A thousand of them send 600,000 requests a month.
The prototype math
A prototype request is small. A short system prompt, the question, maybe one document. Call it 2,000 input tokens and 300 output.
- One request: 2,000 × $3/M + 300 × $15/M, about $0.0105.
- 600,000 requests: about $6,300 a month.
That's the figure that goes into the plan. Here's what it leaves out.
Context grows until someone stops it
In production, the system prompt collects every rule the team added after a bad answer. Tool definitions ride along on every call. Retrieval adds five or six chunks. Chat history comes back with every turn, so the fifth message in a conversation carries the first four. A1's typical production request was 8,000 input tokens and 500 output.
- One request: 8,000 × $3/M + 500 × $15/M, about $0.0315.
- 600,000 requests: about $18,900 a month.
That's three times the prototype figure, and nothing new has shipped. Look at where it goes: $0.024 of the $0.0315 is input. Most of what you pay for is text the model reads, not text it writes. That's why trimming context is usually the first lever.
One question, several model calls
When the feature uses tools (look up an order, check a calendar), one user question becomes several model calls. The model asks for the tool, your code runs it, and a second call sends the whole context back with the tool result added. An agent loop does this many times for a single question.
Say one request in four triggers one tool call. That's 150,000 extra model calls a month at roughly the same size, about $4,700.
Price your own AI feature
Give ChatGPT your token counts and usage and get a monthly cost per active user, including the lines a prototype leaves out.
“|”
Retries are billed too
Timeouts, rate limits, malformed JSON and empty answers all end in a retry, and a failed call that produced tokens still bills them. A1 assumed 10%. On the roughly $23,600 above, that's about $2,400.
Work nobody asked for in the chat
The beta has no background jobs. Production grows them quickly: a nightly summary per user, re-embedding documents when they change, tagging every new ticket, enriching every new lead. They never show up in a demo because nobody watches them run.
Say every active user gets a nightly digest that reads 20,000 tokens of the day's activity and writes 1,000. That's $0.075 per user per night, or $2,250 a month.
Evals cost money, and they're worth it
If you took A1's advice, 50 to 200 real questions run on every prompt or model change. Take 200 questions, one call to answer and one to grade, about 100 runs a month at the same request size: roughly $1,260. It's the cheapest insurance on the list.
The production math
| Line (illustrative) | Monthly |
|---|---|
| Interactive requests, 600,000 at $0.0315 | $18,900 |
| Tool steps, 150,000 extra calls | $4,700 |
| Retries, 10% | $2,400 |
| Nightly digests, 1,000 users | $2,250 |
| Evals, 100 runs of 200 questions | $1,260 |
| Total | about $29,500 |
About $29.50 per active user per month, against a prototype estimate of $6.30. And that's model calls only. Embeddings, the vector database, tracing and any paid tools come on top. If the feature searches the web through Anthropic's search tool, for example, that's $10 per 1,000 searches before tokens.
Averages also hide your best customers. A user who sends 100 requests a day instead of 20 costs about $95 a month in interactive calls alone. On a $29 seat, your most engaged customer is your least profitable one.
Setting a price for AI credits or an AI add-on?
We build the cost model from your production logs, per active customer, then fix what makes it grow. Flat price, one senior engineer, your repo and your cloud account. Send your email and we'll set up a 30-minute call within 24 hours.
Five levers, in the order we'd pull them
Each lever is separate math on the same illustrative model. They overlap, so don't add them up.
- Trim context. Cap chat history, retrieve fewer and better chunks, shorten tool descriptions, delete rules nobody can explain. Going from 8,000 to 3,000 input tokens takes a request from $0.0315 to $0.0165.
- Cache the stable prefix. Put the system prompt and tool definitions first and mark them for caching. Anthropic bills cache reads at 10% of the base input price (pricing). If 2,000 tokens of every request are identical, that saves about $0.005 a request, roughly $3,000 a month here. Caching whole answers to repeated questions is a separate lever that our RAG post puts at 30 to 50% of model spend.
- Route easy requests to a smaller model. Claude Haiku 4.5 lists at $1 and $5 per million tokens, a third of our example price. Route 40% of requests there without failing your evals and the interactive line drops by about $5,000.
- Batch the background work. A nightly digest can wait. Anthropic's Batch API is 50% cheaper, which halves the digest line to about $1,125.
- Meter the heavy users. A daily cap or paid credits above fair use turn the $95 customer into a pricing decision.
One warning about model upgrades. Anthropic says its Claude 4.7 and later models use a newer tokenizer that produces about 30% more tokens for the same text. A lower price per token and a higher token count can cancel out, so measure on your own traffic before you switch.
The cost check
- Do you know average input and output tokens per request from production logs, not from the demo?
- Do you know how many model calls one user question makes, tool steps included?
- Is every background job that calls a model listed, with its schedule?
- Are retries and eval runs in your monthly cost?
- Do you know what your heaviest 10% of users cost, not only the average?
- Is the stable part of your prompt cached?
Related reading
- Your AI prototype works. Here is why it will fail in production: the six gaps, including the cost gap this post expands
- RAG done right in 2026: hybrid search, reranking, evals: evals and answer caching in detail
- MVP cost in 2026: real numbers from $5K to $50K: what the build costs before the model bill starts
The service: AI Product Development. For founders building on AI: Start Matter for AI product founders.
Enjoyed this article? Share it with others
Related Posts

The AI Feature Scorecard: Which Idea Should Your SaaS Build First?
Every SaaS has an AI backlog now, and most of it was ranked by how good the demo looked. Score each idea on five lines instead: how often the job happens, whether the data is ready, whether a right answer can be checked, what a wrong one costs, and which number it moves.

Shopify or WooCommerce With AI vs Headless Commerce With AI
Headless commerce won't make your AI smarter. What it changes is who can reach your data and your rules. If the AI you want already lives inside your platform, stay. If it needs your own data across several systems, the platform becomes the limit.

AI Agents Are Making API Integration More Important, Not Less
Agents don't make APIs obsolete. They make them the part of your product that matters most. The model decides what to do; your API decides what's allowed and does it. Where MCP fits, what to expose first, and seven checks before an agent calls you.