The AI Pricing Inversion: Why the Winners Won’t Sell Tokens

Token prices fell 200x in two years. The cost to train a frontier model collapsed from $100 million to $5 million. And yet the companies capturing the most value from AI aren’t the ones selling compute at the lowest price. They’re the ones who stopped selling compute altogether.

The AI industry is about to go through a pricing inversion that most founders and investors are not prepared for. Everyone is watching the cost curve. Almost nobody is watching the value curve. And the gap between the two is where the next wave of asymmetric returns will come from.

Here’s the thesis: The entire AI pricing model will get repriced around the value of the decision it enables, not the cost of compute behind it. The companies that figure this out before the market commoditizes tokens will capture margins that look nothing like software - they’ll look like advisory, insurance, or legal services.

The evidence is already showing up in the data, and it’s moving fast.

Seat-based pricing, the model that powered two decades of SaaS growth, dropped from 21% to 15% of B2B software companies in just twelve months, according to Growth Unhinged’s 2025 State of Monetization report. Hybrid models surged from 27% to 41% in the same period. IDC now forecasts that by 2028, 70% of software vendors will have refactored their pricing away from per-seat models entirely.

The reason is structural, not cyclical. When AI replaces the work a seat used to do, charging per seat becomes self-defeating. If your product succeeds — if it actually automates what it promises — your customer needs fewer of the thing you charge for. That’s not a pricing problem. That’s a business model contradiction.

And compute-based pricing isn’t the answer either. Token costs are collapsing. The same query that cost $10 per million output tokens on GPT-4 at launch now runs for under $0.60 on GPT-4o Mini — a 94% decline in roughly 18 months. DeepSeek offers comparable capability at $0.30 per million tokens. When the input commodity deflates this fast, any business model anchored to its cost gets dragged down with it.

This is where most analysis stops. Costs are falling, margins are compressing, competition intensifies. The standard technology commoditization story.

But that misses the real shift.

The companies pulling ahead aren’t optimizing on the cost side. They’re redefining the unit of value.

Look at what Intercom did. In 2023, they abandoned per-seat pricing for their AI agent, Fin, and moved to $0.99 per resolution — charged only when the AI actually solves a customer’s problem. Not per conversation. Not per token. Per outcome.

The results: Fin went from $1 million to over $100 million in ARR in under two years. Resolution rates climbed from 27% at launch to over 56%. Intercom even introduced a $1 million performance guarantee — if Fin doesn’t hit agreed resolution targets, they reimburse the customer.

This isn’t a pricing experiment. It’s a structural repositioning. And the logic behind it reveals something important about where AI value capture is headed.

Intercom’s own head of pricing explained the reasoning clearly: if Fin works as designed, customers will need fewer human support agents over time. Seat-based pricing would have meant Intercom’s revenue shrinks as its product improves. Outcome-based pricing means revenue grows with performance.

That alignment between product performance and revenue is the critical insight. And it’s the same alignment that has defined pricing in every regulated, high-stakes industry for decades.

Think about how advisory, insurance, and legal services work.

A management consulting firm doesn’t charge by the hour of compute time their analysts spend on your problem. They charge based on the complexity of the decision, the stakes involved, and the value of getting it right. The fees for a pricing strategy engagement isn’t determined by the number of spreadsheets produced. It’s determined by the revenue at risk.

Insurance premiums aren’t priced on the cost of running actuarial models. They’re priced on the risk transferred. A $10 million directors and officers policy costs what it costs because of what it covers, not because of the computational effort behind the risk assessment.

Legal billing increasingly works the same way. Contingency fees, success fees, and value-based arrangements are displacing billable hours precisely because clients want to pay for outcomes — for the merger that closes, the case that settles, the regulatory approval that lands.

In all three cases, the pricing model is anchored to the value of the decision enabled, not the cost of the input consumed. The input cost is almost irrelevant to the buyer. What matters is: did it work? How much was at stake? What’s the alternative?

AI is heading to exactly the same place. And the transition will be faster than most expect, because the cost side is deflating so rapidly that it forces the question. When the token costs essentially nothing, the only defensible pricing is based on what the token produced.

This creates three distinct strategic implications.

For founders building AI products: The window to establish value-based pricing is now, not after commoditization forces it. Companies that wait until token costs reach near-zero to rethink their pricing will find themselves in a race to the bottom with no differentiation. The companies that define their value metric early — per resolution, per contract reviewed, per lead qualified, per decision supported — build customer relationships anchored to outcomes. That’s dramatically harder to displace than a relationship anchored to a cost-plus markup on API calls.

The operational challenge is real. Value-based pricing requires you to measure outcomes, not just usage. It requires you to absorb cost variability. It requires sales teams that can articulate ROI, not just features. But the companies doing this are already seeing 2-3x higher customer traction than those monetizing AI as a bundled feature.

For investors evaluating AI companies: Pricing model is now a leading indicator of durability. AI companies averaging 50-60% gross margins — the current industry norm — are not SaaS businesses, no matter what their pitch deck says. Traditional SaaS runs at 80-90% gross margins. The companies that will close that gap are the ones shifting from cost-plus to value-based pricing, where the margin expansion comes from the spread between what the AI costs to run and what the outcome is worth to the buyer.

That spread is the real asset. A token costs fractions of a cent. A resolved customer ticket saves $5-15 in human agent cost. A reviewed contract saves $200-500 in associate billing time. A qualified lead is worth $50-500 depending on the sales cycle. The companies that price against the value of the outcome — not the cost of the inference — are the ones building durable margin structures.

For operators in incumbent companies: Watch how your software vendors are repricing. AI add-ons are already adding 30-110% to base SaaS costs. Microsoft Copilot represents a 60-70% premium over base licensing. If your vendor is charging you for AI based on seats or tokens, you’re paying for inputs. If they start charging per outcome, the negotiation changes entirely — and your leverage shifts from volume discounts to performance guarantees.

The AI pricing inversion isn’t a trend to monitor. It’s a structural repricing that separates the companies building commodities from the ones building moats.

Every major technology cycle goes through this. Hardware commoditizes, so value moves to software. Software commoditizes, so value moves to platforms. Compute commoditizes — and we’re watching it happen in real time — so value moves to outcomes. The companies that price on the outcome side of this equation will compound. The ones stuck on the cost side will compress. Capital will follow accordingly.

𝘈𝘯𝘺 𝘷𝘪𝘦𝘸𝘴 𝘰𝘳 𝘴𝘵𝘢𝘵𝘦𝘮𝘦𝘯𝘵𝘴 𝘦𝘹𝘱𝘳𝘦𝘴𝘴𝘦𝘥 𝘢𝘳𝘦 𝘮𝘪𝘯𝘦 𝘢𝘯𝘥 𝘯𝘰𝘵 𝘵𝘩𝘰𝘴𝘦 𝘰𝘧 𝘮𝘺 𝘦𝘮𝘱𝘭𝘰𝘺𝘦𝘳