The assumptions
A “user cost” is just calls per month times cost per call, and cost per call is set by token counts and the model’s price. So we define three usage profiles and one representative call size for each, then run every profile across five models. Adjust any of these to your own product; the method is the point, not the exact inputs.
Profile Calls / month Tokens per call (input / output)
--------------------------------------------------------------------
Light user 100 500 in / 300 out (short chat turns)
Heavy user 1,500 500 in / 300 out (daily power user)
Agent user 750 2,000 in / 500 out (50 runs x ~15 calls,
large context per step)Prices are published token prices per million tokens (input / output) as of July 2026. Prices change often, so verify before you rely on any number. Sources: OpenAI pricing and Anthropic pricing. For the full cross-provider table including DeepSeek and Gemini, see DeepSeek vs OpenAI vs Gemini cost.
Model Input $/M Output $/M
-----------------------------------------------
GPT-4o-mini 0.15 0.60
Gemini 2.5 Flash 0.30 2.50
GPT-5 ~1.25 10.00
GPT-4o 2.50 10.00
Claude Sonnet 4.6 3.00 15.00Cost per user per month, modeled
Running each profile across each model gives the table below. Cost per call is (input_tokens × input_price + output_tokens × output_price) / 1,000,000, then multiplied by calls per month. These are rounded.
Model Light (100) Heavy (1,500) Agent (750)
----------------------------------------------------------------
GPT-4o-mini $0.03 $0.38 $0.45
Gemini 2.5 Flash $0.09 $1.35 $1.39
GPT-5 $0.36 $5.44 $5.63
GPT-4o $0.43 $6.38 $7.50
Claude Sonnet 4.6 $0.60 $9.00 $10.13Two patterns jump out. First, the model choice moves the cost per user by roughly 20x across this table, for the same user doing the same work. Second, the agent profile is not much heavier than the heavy-chat profile on a cheap model, but on a frontier model it is the most expensive column, because agents spend big context on every one of their many internal steps. And this assumes the agent behaves. A single reasoning loop that fans out 10x turns that $10 into $100.
The margin math: same price, opposite outcomes
Cost only means something next to revenue. Put every modeled user on the same $29/month plan and compute margin (plan revenue minus AI cost):
On a $29/month plan Cost Gross margin
------------------------------------------------------
Light user, GPT-4o-mini $0.03 ~99.9% very profitable
Heavy user, GPT-4o-mini $0.38 ~98.7% very profitable
Heavy user, Claude Sonnet $9.00 ~69% healthy, watch it
Agent user, Claude Sonnet $10.13 ~65% healthy, fragile
Agent user, looping (10x) ~$101 NEGATIVE loses ~$72/userThis is the whole thesis of AI unit economics in one table. The subscription price is fixed, but the cost is not, so the same $29 plan yields a 99.9 percent margin on one customer and a negative margin on another. The averages hide it: blend these users together and you might report a comfortable 85 percent gross margin while a handful of accounts quietly lose you money every month. That is exactly the customer you cannot see from the provider bill, and exactly the one worth finding.
Why your real number will differ (and why that is the point)
Treat this benchmark as a map, not a measurement. Several things move your real cost per user away from the table:
Prompt size
We assumed modest prompts. If you stuff a long system prompt, few-shot examples, or a big RAG context into every call, your input tokens can be 5x to 20x ours, and input cost scales with them. Prompt caching, where a provider discounts repeated context, can pull it back down. Neither is in this model.
Model mix
Almost no real product runs one model. You might route cheap features to GPT-4o-mini and reserve a frontier model for the hard ones, which lands you somewhere between the columns. Deciding that split well is its own exercise: see GPT-5 vs GPT-4o-mini for how to route by a blind quality test.
Usage distribution
The biggest unknown is who your users actually are. Real AI usage follows a long tail: most users look like the light column, a few look like the agent column, and the few are where the losses concentrate. A single average cost per user tells you almost nothing; you need the distribution, which is why you calculate margin per user rather than in aggregate.
How to measure your own instead of trusting a table
The honest end of any benchmark is: go get your real number. That means attributing cost to a userId on every call, recomputing dollars from current pricing so it never drifts, and aggregating per user across a full month. You can build that, or wrap your existing OpenAI, Anthropic, or Gemini client with Weckr and get per-user cost and margin as a byproduct of the calls you already make, with per-plan spending caps to bound the agent-looping row above.
See the per-user view on seeded data, no signup, at useweckr.com/demo, and model your own plan against these costs with the break-even calculator. This benchmark is one section of the larger guide to AI cost and margin management.
FAQ
How much does one AI user cost per month?
It depends entirely on the model and how heavily they use the feature, and the range is enormous. Modeled from public 2026 pricing, a light assistant user (about 100 calls a month) costs roughly 3 cents on GPT-4o-mini but around 60 cents on Claude Sonnet. A heavy user (about 1,500 calls) costs roughly 38 cents on GPT-4o-mini but about $9 on Claude Sonnet. An agent user whose runs fan out into hundreds of calls can cost $10 or more, and far more if a run loops. The same subscription price can therefore produce a 90 percent margin on one user and a negative margin on another.
Is this benchmark based on real user data?
No. These figures are modeled from published provider token prices as of July 2026 and a set of stated usage assumptions (calls per month and tokens per call). They are meant to show the shape and the spread, not to be your exact numbers. Your real cost per user depends on your prompt sizes, your model mix, caching, and your actual usage distribution, which is why the honest recommendation is to measure your own rather than trust any benchmark, including this one.
Why is the cost per user so different between models?
Because token prices span more than 200x from the cheapest flash-tier models to frontier models on input, and roughly 10x to 25x on output. Output tokens usually dominate a real bill, so a model that charges $15 per million output tokens costs about 25x more than one at $0.60 for the same generated text. Multiply that by a heavy user’s call volume and the per-user gap becomes the difference between profitable and underwater.
How do I find my own real cost per user?
Attribute cost to a userId on every LLM call, recompute the dollar cost from current provider pricing rather than trusting a static table, and aggregate per user over a full month. You can build that yourself, or wrap your existing OpenAI, Anthropic, or Gemini client with Weckr and get per-user cost and margin as a byproduct of the calls you already make. See it on seeded data at useweckr.com/demo.
The number that is not on this page
The most important cost per user in this article is the one we could not compute for you, because it depends on your prompts, your model mix, and your users. The table shows you the spread and where the danger is: heavy usage on an expensive model, and agents that loop. Find your own version of this table on your own data at useweckr.com/demo, and you will know which customers to price up, cap, or celebrate.