LLM API cost calculator: what an AI feature costs per user per month on OpenAI, Claude and Gemini

How much does an LLM API cost per user?

An LLM API costs between a few cents and a few dollars per active user per month for most product features. The range is wide because three things multiply: how often a user triggers the feature, how many tokens each request sends and receives, and which model answers. The calculator above does that multiplication for nine current models from OpenAI, Anthropic and Google.

Provider pricing pages quote prices per million tokens. That is useful for comparing models and nearly useless for a budget, because nobody plans a product in tokens. A founder or product manager needs an answer like "the support assistant will cost $0.55 per user per month on a mid-tier model, or $0.03 on a small one". That is what the calculator shows: cost per feature, per month and per user.

All prices below were checked on the official pricing pages on October 22, 2026. Prices change several times a year, sometimes with promotional periods: Google has already announced that Gemini 4 Argon doubles in price on December 1, 2026, and Gemini 3.8 Flash on January 1, 2027. Tick the checkbox in the calculator to see the bill at the announced prices.

Which models does the calculator compare?

The calculator covers the flagship, mid-tier and small model of each provider. Prices are in USD per million tokens for the standard API tier and normal prompt lengths, as of October 22, 2026.

A few details the headline prices hide:

• Claude Haiku 5.5 costs five times more when a single prompt is over 100,000 tokens. The calculator switches to that price automatically. OpenAI's long-context prices start above 272,000 tokens, which the calculator does not reach.
• OpenAI and Anthropic charge 1.25x the input price the first time a prompt prefix is written to the cache. On a busy feature that fee is spread across thousands of hits, so the calculator leaves it out.
• Gemini's explicit context caching adds a storage fee per hour. Implicit caching, which is what most apps use, does not.
• Mid-tier prices are now very close across providers. The difference between providers on one task is usually smaller than the difference between tiers.

How do you estimate the cost of an AI feature?

You estimate the cost of an AI feature by counting one typical request, then scaling it by traffic. This is the same order we follow when we scope an AI feature for a client.

1. Write down one real request. Take the actual system prompt, the tools, the documents you attach and a typical user message. Count the tokens with the provider's token counter. Teams usually underestimate this number by 2-3x, because the system prompt, tool definitions and chat history grow quietly.
2. Measure a typical answer. Count output tokens on 20-30 real answers, including reasoning tokens if the model thinks before answering. Set a max_tokens limit you are happy to pay for.
3. Estimate how often users trigger it. Requests per active user per day, from analytics on a similar feature or a pilot. A chat turn, a search, a generated document and a background job each count as one request. Agents that call tools count every model call, often 3-10 per task.
4. Decide what can be cached. Everything that repeats at the start of the prompt: instructions, tool definitions, a product catalog, a long document the user asks several questions about.
5. Decide what can wait. Anything not shown to a user in real time can go through the Batch API at about half price.
6. Multiply and add a margin. Retries, failed calls, evaluation runs and prompt experiments usually add 10-30% on top.

The calculator does steps 3-6 for you. Steps 1 and 2 are where the accuracy comes from, so after launch replace the preset numbers with the averages from your own logs.

Worked examples: three features for 1,000 active users

The three examples below use the calculator's presets and 1,000 active users a month. The token counts are typical ranges we see in first versions of these features, not measurements from one specific project.

Support chatbot

Two conversation turns per user per day, 6,000 input tokens per turn (system prompt, three help articles, chat history) and 400 output tokens. About 60% of the input is the same on every turn and is cached. That is 60,000 requests, 360 million input tokens and 24 million output tokens a month.

Here is the arithmetic for GPT-6.1 Sol, so you can check the calculator by hand: 144 million uncached input tokens x $2 = $288, plus 216 million cached tokens x $0.10 = $21.60, plus 24 million output tokens x $10 = $240. Total $549.60, or $0.55 per user. Without caching the same feature costs $960.

The spread between the cheapest and the most expensive model is almost 100x. Most support bots work well on a small or mid-tier model, so this is where testing quality on your own conversations pays off the most.

Document search over your files (RAG)

Five questions per user per day. Each question sends 8,000 tokens of retrieved document chunks and instructions, and gets a 500-token answer. Only 30% of the input is cached, because the retrieved chunks change with every question. That is 150,000 requests and 1.2 billion input tokens a month.

On a mid-tier model this comes to about $2,466 a month, or $2.47 per user. On Gemini 3.8 Flash it is $938, and on the small OpenAI and Claude models $125. The input side dominates, so the most effective savings are retrieving fewer, better chunks and ranking them before they reach the model. Our RAG implementation guide covers how to do that without losing answer quality.

Resume or long text generation

One document per user per day: 3,000 tokens of instructions and user data in, 1,500 tokens of finished text out. This is the shape of features like the resume and cover letter writer in our own product, AI Resume Master, which reached 50,000 monthly active users on the OpenAI API.

On a mid-tier model the bill is about $545 a month, or $0.54 per user, and $450 of it is output. Output makes up 83% of the cost on this preset, so caching barely matters here. Shorter outputs, a smaller model for drafts, or generating sections only when the user asks for them save far more.

How much do prompt caching and the Batch API save?

Caching and batch are the two discounts every provider offers, and they stack.

Prompt caching bills the repeated start of a prompt at 2.5-10% of the normal input price, depending on the model. In the chatbot example caching cuts the mid-tier bill from $960 to $550, a 43% saving. It only works when the cached part is byte-for-byte identical and comes first, so order matters: system prompt, tool definitions and fixed documents at the top, then the conversation, then the new message. A timestamp or user name at the start of the system prompt breaks the cache on every request.

The Batch API processes requests asynchronously, with results within 24 hours, at about half price on input and output. For the classification preset (20 short calls per user per day, 600,000 requests a month) batch cuts a mid-tier bill from $1,503 to $752. It fits ticket tagging, extracting fields from documents, enriching CRM records, nightly reports and evaluation runs. It does not fit anything a user waits for.

Together they matter most for high-volume background work. A classification job that runs nightly with a cached instruction block pays a fraction of what the same job costs as live, uncached calls.

How can you reduce LLM API costs without hurting quality?

These are the changes we make most often when a client's AI bill grows faster than its user base, roughly in order of savings per hour of work:

1. Route by difficulty. Send most requests to a small model and escalate to a mid-tier or flagship model only when the request is hard or the small model is unsure. On the support bot above, routing 80% of turns to a small model brings the bill from $550 to around $135.
2. Cache the stable part of every prompt. Usually a one-day change that cuts input cost by a third or more on chat features.
3. Trim the context. Send the three most relevant chunks instead of ten, summarize long chat histories, and drop tool definitions the current step does not need.
4. Cap and shape output. Ask for structured output with only the fields you need, and set max_tokens. Output is the expensive side.
5. Move background work to batch. Anything nobody watches live.
6. Set limits per user. A rate limit per user and per tenant stops one heavy account or a script from turning into a surprise invoice.

Measure quality before and after each change on 50-100 real examples. A cheaper model that gets 5% more answers wrong can cost more in support tickets than it saves on tokens.

Should you choose a provider by price?

Not on price alone. At the mid tier, OpenAI, Anthropic and Google now charge almost the same per token, and the cheapest model on paper is not always the cheapest per finished task. A model that needs two retries or a longer prompt to get the answer right costs more per task than a pricier model that gets it right the first time. Tokenizers also differ: the same text turns into a different number of tokens on each provider, and Anthropic notes that its newer models produce about 30% more tokens for the same text than its older ones.

The reliable method is to run your top two or three candidates on the same 50-100 real requests and compare cost per correct answer. Our Claude API vs OpenAI API comparison covers the differences that matter beyond price, such as tool use, structured output and data terms.

Need help putting an AI feature into production?

We build AI features into existing products and measure their cost per user before launch, so the monthly bill is a known number. See our AI integration services, or the provider-specific pages for OpenAI API integration and Claude API integration. Projects start from $10,000 with a fixed quote after a short scoping call, and a first AI feature often fits into one AI Sprint: $10,000 fixed for 4 weeks.