Pricing AI Products When the Cost per Task Keeps Falling

Every AI company is building on an input that gets cheaper every quarter, and most of them are still pricing like it won't. This month alone, another frontier lab cut the price of its smallest model sharply. Nobody was surprised. That's the part worth paying attention to. A falling cost per task isn't news anymore, it's the weather.

The interesting question isn't whether the cost goes down. It's who keeps the difference.

the cost curve only goes one way

A task that cost a rupee to run last year costs a fraction of that today, and will cost a fraction of that again next year. Model providers compete on price, open weights models keep closing the gap, and inference keeps getting more efficient.

There's a catch. Cheaper tokens don't always mean a cheaper task. Agents think longer, call more tools, retry more often, and check their own work. The cost per token falls while the tokens per task climb. So the bill doesn't shrink as neatly as the price sheet suggests, and it moves in both directions depending on how the product is built.

That makes pricing harder, not easier. The input cost is falling and unpredictable at the same time.

seat pricing breaks first

Seat pricing worked because software was a tool and every tool needed a person to use it. Ten people, ten seats.

AI products increasingly do the work instead of helping someone do it. If an agent handles what three support reps used to handle, the customer doesn't want three seats. They want fewer. The better the product gets, the fewer seats it sells. That's a strange business to be in.

Some companies are patching this by charging per agent, as if an agent were a seat. It buys time. It doesn't fix the underlying problem, which is that the unit being priced has nothing to do with the value being delivered.

usage pricing hands the discount to the customer

Usage pricing feels like the honest answer. You pay for what you use. Costs go down, prices go down, everyone's happy.

Except the company has now tied its revenue directly to the one number guaranteed to keep falling. Every time the underlying model gets cheaper, a usage priced product either cuts its price or watches a competitor do it first. The savings flow straight through to the customer. The company is left with the same margin on a smaller number.

It's a bit like a Bangalore auto on the meter. Fair, transparent, and the driver earns exactly what the distance is worth, no matter how good he is at finding the shortcut. If anything, the shortcut costs him.

Usage pricing works for infrastructure, where the customer is technical and the job is literally to resell compute. For an application, it turns the product into a pass through for someone else's price cuts.

outcome pricing is where the margin lives

The alternative is pricing the work itself. Per resolved ticket. Per contract reviewed. Per qualified lead. Per line of code merged.

Here the falling cost works in the company's favour. The customer pays for a resolved ticket whether it took a thousand tokens or a hundred. When the model gets cheaper, the price stays where it is and the margin widens. That's the only pricing model where the cost curve is a tailwind instead of a threat.

It's the fixed fare on a cab app. The customer knows the price upfront and doesn't care about the route. The operator who finds the faster route keeps the difference.

Outcome pricing is also the hardest to get right. Three problems keep coming up.

1. Attribution. If a deal closes, how much of it was the AI and how much was the salesperson? Customers will argue this every renewal.
2. Definition. "Resolved" has to mean the same thing to both sides. A ticket closed by the agent and reopened the next morning isn't resolved.
3. Variance. Some tasks are cheap and some are expensive, and a single price per outcome has to absorb both without the company losing money on the hard ones.

Companies that solve these get a pricing model that improves on its own. Companies that don't end up negotiating every invoice.

what Jio already taught India about this

When Jio launched in 2016, mobile data in India went from expensive to almost free within a couple of years. Consumption exploded. The operators that survived stopped selling data by the megabyte and started selling plans, a fixed price for a bundle of what people actually wanted, calls, data, validity.

AI is going through the same thing, just faster. Pricing by the token is pricing by the megabyte. It made sense while the input was scarce. Once it isn't, the customer stops caring about the unit and starts caring about the outcome, and the company that prices the outcome captures the value that the falling cost creates.

what this means when looking at a company

When I look at an AI company now, pricing tells me more than the demo does. A few things I check.

• Does the price move with the model, or with the value? If every model price cut forces a price cut, the company is a reseller with a nicer interface.
• Which way is gross margin heading? For a well priced AI product, margin should improve as models get cheaper. Flat or falling margin in a falling cost environment is a pricing problem, not a cost problem.
• Can the company switch models without the customer noticing? The ability to route work to whichever model is cheapest for the task is where the margin actually comes from. That only pays off if the price isn't tied to the model.
• Is the unit something the customer already counts? Tickets, contracts, hires, invoices. If the customer has to learn a new unit to understand the bill, the bill will get questioned.

conclusion

The cost of intelligence will keep falling. That part is settled. What isn't settled is whether that saving shows up as lower prices for customers or as margin for the companies building the products.

The answer depends almost entirely on what the company chose to put a price on. Price the tokens and the savings belong to the customer. Price the work and they belong to you.