Pricing an AI Product When Your COGS Is a Token Bill

If your product calls a model, your cost of goods sold moves with usage, and that breaks the pricing instinct most founders inherited from subscription software. The asymmetry is sharper than people expect: across both major providers' published list prices, output tokens cost roughly five times input tokens. Anthropic lists Sonnet 5.5 at $2 per million input tokens and $10 per million output, and OpenAI lists gpt-6.1-sol at the same $2 and $10. Your most engaged users generate the most output, so under flat per-seat pricing your best customers are also your least profitable.

That is the whole problem in one sentence. What follows is how to price around it.

Why does flat per-seat pricing fail for AI products?

Because in a traditional subscription business, serving one more active user costs close to nothing, so engagement and margin point the same way. When inference is in the loop, they point in opposite directions.

Picture two accounts on the same €50 per month plan. One runs a handful of short queries a week. The other has wired your product into a daily workflow and generates long outputs continuously. In a conventional software business, the second account is the success story. With a token bill attached, the second account can cost more to serve than it pays, while the first subsidises it.

The failure is slow and therefore dangerous. It does not show up in aggregate gross margin while most accounts are light. It shows up later, as adoption deepens, which is exactly when the board deck says things are going well.

How do you calculate gross margin on an AI feature?

Per unit of work, not per customer, and starting from the token counts rather than the invoice.

For each core action in your product, measure four things: average input tokens, average output tokens, the proportion of input that can be served from cache, and how many times a typical account performs that action per month. Then price it against the provider's published rates.

A worked illustration using Anthropic's published Sonnet 5.5 rates of $2 per million input and $10 per million output. Suppose one action sends 4,000 input tokens and returns 1,000 output tokens. Input costs 4,000 / 1,000,000 × $2, which is $0.008. Output costs 1,000 / 1,000,000 × $10, which is $0.01. Total around $0.018 per action, and note that the output, at a quarter of the token volume, is more than half the cost. An account running 500 of those a month costs roughly $9 to serve. On a €50 plan that is comfortable. At 2,000 actions a month it is $36, and the plan is no longer comfortable at all.

Do this arithmetic for your three most common actions before you set a price, not after. The numbers are public, so there is no excuse for guessing.

What are the actual levers on inference cost?

Four, in rough order of how much they matter.

Output length. Since output is the expensive half, every unnecessary sentence your model generates is a margin decision. Constraining response length, returning structured data instead of prose where the interface allows it, and resisting the temptation to pad answers all show up directly in COGS.

Caching. Both providers charge dramatically less for input served from cache. OpenAI's published pricing shows cached input at 90% below standard input for its current model families. Anthropic publishes separate cache read and write rates, with cache reads priced far below base input and cache writes priced above it, on a five-minute window. If your product sends the same long system prompt or document context repeatedly, caching is usually the single largest saving available, and it is an architecture decision rather than a negotiation.

Batching. Both providers publish a 50% discount for batch processing. Anything that does not need to be synchronous, overnight enrichment, bulk classification, scheduled summaries, belongs in a batch queue. Many teams run asynchronous work through the real-time path purely out of habit.

Model routing. Published prices vary by more than an order of magnitude across a single provider's range. Anthropic lists Haiku 5.5 at $0.10 per million input tokens for prompts up to 100K, against $4 for Opus 5.5. Routing straightforward work to a smaller model and reserving the expensive one for genuinely hard requests is normally worth more than any prompt micro-optimisation.

These are published list prices at the time of writing and they move, often downward. Build your model so the inputs can be updated, and re-run it when they change.

Which pricing model should an AI-native product use?

There are four workable shapes, and the right answer depends on how predictable your usage is.

Per seat with a usage ceiling. Familiar to buyers, and the ceiling caps your downside. The honest version states the limit plainly rather than burying it in a fair-use clause nobody reads.

Usage-based. Margin follows cost automatically, which is the cleanest economics available. The cost is buyer discomfort: procurement departments dislike unpredictable invoices, and an unbounded bill slows enterprise deals.

Hybrid. A platform fee covering access plus metered consumption above an included allowance. This is where most AI-native companies we see end up, because it gives the buyer a predictable floor and gives you cost recovery at the top.

Outcome-based. Charging per resolved ticket, per completed document, per qualified lead. Commercially the most compelling when your product genuinely replaces a unit of work, and operationally the hardest, because you need an agreed definition of success and the data to prove it.

The mistake to avoid is picking a model from what competitors display on their pricing page. You cannot see their token economics, their caching architecture or their discounts, so you are copying a conclusion without the premises. Our general guide to startup pricing strategy covers the commercial side of that decision.

What do investors ask about AI unit economics?

Four questions, and we ask them early because the answers predict a lot.

What does your most expensive account cost to serve, and what does it pay? Not the average. The outlier, because the outlier is where everyone else is heading as the product gets better.

What is gross margin on the core action, and how has it moved over the last six months? The direction matters more than the level. A margin improving as volume grows means the architecture is learning. A margin eroding as engagement deepens means pricing and cost are pointed in opposite directions.

What happens to your economics if model prices fall by half? Usually a pleasant answer, and worth having ready.

And if your usage per account doubles? This one separates teams who have modelled their costs from teams who have read their invoice. If doubling engagement makes the business worse, that is not a pricing detail, it is the business model.

Founders sometimes expect us to penalise lower gross margins than a pure software business would show. We do not, as long as the number is understood, deliberate and moving in the right direction. What does concern us is a founder who cannot produce it. This connects directly to our thesis on AI-first company building, and it belongs in the same model as your runway calculation, because inference cost is one of the few operating expenses that scales with success. If you are preparing to raise, our guide to early-stage VC firms in Europe covers who invests at this stage.

Frequently asked questions

Are output tokens really more expensive than input tokens? Yes, and by a wide margin. Both major providers' published pricing shows output at roughly five times input for comparable models, for example $2 against $10 per million tokens for Anthropic's Sonnet 5.5 and OpenAI's gpt-6.1-sol. Response length is therefore a direct cost lever.

How much does prompt caching save? It depends on the provider and model, but the effect is large. OpenAI publishes cached input at 90% below standard input for its current families, and Anthropic publishes separate cache read rates well below base input. If you repeatedly send the same context, caching is usually the biggest single saving available.

Should an AI product use usage-based pricing? It gives the cleanest margin alignment, because revenue moves with cost. Most companies we see settle on a hybrid instead, with a platform fee plus metered usage above an allowance, because buyers want a predictable floor and sellers want cost recovery at the top.

What gross margin should an AI-native company target? There is no single right number, and treating a figure from a blog as a benchmark is a mistake. What matters is that you can calculate your margin on the core action, explain the drivers, and show it moving in the right direction as volume grows.

Does falling model pricing solve the problem on its own? It helps, and published prices have generally moved downward. But usage per account tends to rise as products get better, so cheaper tokens often get spent rather than saved. Price for the economics you have today, and treat future price drops as upside rather than plan.