AI Pricing Thoughts
Fair to your customers, sustainable for your business; a practitioner's playbook for pricing when AI usage is variable
Over a year ago I spent too much time thinking about AI pricing models. I was surrounded by folks who weren’t interested in doing math and preferred “we’ll just figure it out” methodologies, which failed, in my opinion.
So, in this read, I wanted to recast my own journey figuring out AI pricing and the journey I’m still on with dozens of other founders.
If you’re reading from the seat of your chair, you likely have a platform or seat based component, and a variable component that depends on using AI in the background. Every time a user runs a feature you burn tokens, and those tokens cost you real money. How do you price the product to be fair to customers while protecting your business?
This one is for founders and executives who are pricing an AI product right now and feeling the tension between charging for value and getting wrecked by variable costs.
I am not going to hand you a magic formula, because there is not one. What I can give you is the way I actually work the problem: the reasons AI pricing is hard, the questions you have to answer before you touch a number, the spreadsheet I rebuild every time, and how to stay transparent and realistic so you do not price yourself into a corner you cannot climb out of later.
This edition of Growth Stack Mafia is brought to you by Blip.
Blip launches Meta ads in bulk from saved templates, so performance teams can spin up and scale campaigns without building each one by hand in Ads Manager.
>> Bulk-launch from your desktop, Google Drive, or Dropbox, with saved templates for copy, CTAs, and UTMs, plus post ID scaling to boost organic Instagram posts.
>> Media-buying analytics with anomaly detection and one-click scale, reduce, or pause recommendations.
I developed this from building pricing three times in my career, most recently as a cofounder building an AI-native CRM.
My first advice to any aspiring founders and executives out there: pricing is a complicated and, ultimately, mathematical process. It is equal parts art and science. The science can take an infinite amount of time and energy, and should be balanced with the art, which is understanding your customer and the environment in which you operate (as a function of time, resources, and other constraints).
Do not skip the math. But also, do not blindly follow the product brute on your team who wants to “move fast and ship.”
The problems
Anytime a product uses AI, it is inherently variably priced, because AI uses tokens, and an action can consume more or fewer tokens depending not only on how it was designed but on how the LLM ends up choosing to execute the action.
>> Features designed to use lower tier models (Sonnet, for example) consume fewer tokens.
>> Features designed to use higher tier models (Opus, Codex) consume more tokens.
>> Beyond the model itself, how the feature is implemented drives consumption: poorly designed prompts consume more tokens; more complex, nuanced, sophisticated prompts consume fewer tokens overall (even if they consume more up front).
>> Which model provider you use makes a difference. In some cases you can degrade or upgrade quality across providers and find similar outcomes. This takes trial and error.
It is quite hard to come up with a “confidence interval” for how much an action or outcome will cost in tokens and price. There is not a deterministic, linear relationship between outcomes and tokens. It is more like a statistical CI. The goal of AI price engineering is to refine how the features and products work over time so that CI becomes reliable.
Most of these problems are hard to solve at once and quickly, especially in an early stage, staff limited organization. Better to work toward these outcomes over time. If you have a lot of smart, mathematically oriented, engineering savvy people, you might get answers faster, but weigh that time against the fact that your product and feature set will keep changing dramatically as you go from $1M to $XXM.
Tldr; better to ship quickly, learn, and adjust.
The questions
Do you charge customers exactly for what you consume? What if you cannot reliably predict the number of tokens an outcome will consume?
Do you let customers bring their own AI LLM API key, so they pay directly for what they consume? That means no margin on AI features.
If you do decide to make margin on AI features, what is fair, and do you have to be transparent about your margin? Not all features and products will make margin. You may elect certain AI products to be loss leaders, subsidized by easy features that consume very little, or use very simple models. If you are in the business of making profit, this is the way.
To be direct: most companies should be trying to make some margin on their AI products.
Why? Because it is not 2020 ZIRP days. Companies are designed to make money and to be sustainable. Some financial misengineering is fine, especially when you are getting started and learning. But if you have priced a product and are making more than $100K ARR, you should be considering how much your products and features cost, and how sustainable they will be in the long run.
If you do not, you could build a highly revenue generating business that is totally unsustainable because of the way you engineered your pricing. The larger the customer base gets, the harder it is to rip the proverbial bandage and change.
Punchline: start somewhere, consider margin, learn, and iterate quickly.
Where to start
Build an AI pricing spreadsheet. It will have many iterations. If you are a technical team, I recommend building this in code, and maybe designing it into an agent that can continually adjust pricing or run pricing analysis on real time data. That way adjusting pricing becomes less of a monthly or quarterly exercise and more of a daily habit. You might not push pricing changes to customers, but you are continuously tracking how variable consumption is driving your business. Financial understanding and engineering is the core of some of the best businesses, including Ramp.
Then, in order:
>> List all your products and features.
>> List what model(s) those products and features use.
>> Create a token consumption CI, or a median or average summary, over whatever time frame you have. If you cannot because you are not tracking tokens, take the time now to have engineering instrument all product surface areas with token consumption using something like Langchain.
>> From token consumption, estimate a per workspace or per user variable cost: the cost of servicing a customer or account based on the AI variable component of their plan.
>> Translate AI tokens into a “credit” or platform token most people can reason about. Example: 1M tokens for 10 enrichments, 10 enrichments = 1 credit. Alternatively, each product or feature may have its own component that you price.
>> Add in your fixed seat or platform fee.
>> Calculate margin without CAC.
If you do not like what you see, you have two options: keep eating the cost on AI variable pricing to drive growth, or modify your pricing plan and adjust the relative weights of each product surface area to hit your margin goals.
Transparency and realism in pricing
Depending on who your customer is and the value you provide, you may want or need to give more visibility into the exact unit economics of your AI features. Savvy, technical buyers who could build your solution themselves will demand detailed pricing understanding.
Take workflows. If your buyers already run them in Zapier, they know the unit economics cold; Zapier sits down around fractions of a cent per run. To compete you have to price at least that aggressively, which can push workflows into loss-leader territory, because a startup cannot buy compute or tokens at hyperscaler scale. Savvy buyers do that math, and at millions of runs the variable cost gets very real.
Never surprise your customers with variable bills.
Amid all this talk of paying for value, most startups cannot command variable, value based pricing. CFOs and financially educated executives are reining back the 2024 to 2025 world of excess AI spending. Only products that deliver exceptional, world class, unprecedented value can charge variable rates that might surprise financial leaders. Be extremely cautious and realistic about how life changing or important your product actually is.
The risk if you are not: the customer decides they can just build the thing you do themselves.
Start somewhere
None of this has to be perfect on day one. The worst move is to freeze because you cannot model every token precisely, or to wave the whole thing away and price on vibes until a big customer base makes the math impossible to change. Start somewhere. Build the spreadsheet, put a rough confidence interval on your token costs, decide where you are willing to run a loss leader and where you need margin, then revisit it on a schedule instead of once a year in a panic.
Pricing in the age of AI is going to keep moving. Model costs are dropping, providers keep leapfrogging each other, and buyer expectations are shifting under all of it. The teams that win will treat pricing as a living system they tune constantly, not a decision they make once and defend forever. Build that habit now, while your customer base is still small enough that you can change your mind.
Additional Reading
>> Austin Hay, Harsh Truths of Sub-$10M ARR Marketing. Why cost discipline matters most before you have scale to hide behind.
>> Austin Hay, A Framework for Evaluating New Opportunities in 2026. How I decide what to build, what to subsidize, and what becomes a loss leader.
>> Kyle Poyar, The state of AI pricing at 40 leading startups. A real look at how AI-native companies actually charge, including credits and hybrid models.
>> a16z, AI Is Driving a Shift Towards Outcome-Based Pricing. The macro case for why per-seat is breaking and variable cost is forcing new models.
>> Lenny Rachitsky, Pricing your SaaS product. A solid grounding in willingness-to-pay methods if you are starting from zero.
This seems to be a cost-based approach. What role does value play in your approach to pricing?