The frontier AI labs are pricing their products in a way that quietly works against them. We see it from inside the companies we work with, week after week, and it has become one of the most consistent patterns in our work. Consumption pricing is suppressing the exact usage the labs need most right now. The fix is two moves: make the model cheap, and make the harness expensive.
This isn't a complaint about cost. It's an observation about behavior, and about timing. There's still a land grab underway for who owns the daily habit of the enterprise, and the current pricing is handing that habit to whoever is willing to subsidize it. What follows is what we're seeing, why we think it's happening, and what we'd do about it.
What we're seeing in the field
Our work puts us inside companies that are trying to actually run on AI rather than experiment with it, and the pattern has become consistent. A company invests real effort getting its people onto a tool like ChatGPT or Claude. Teams get comfortable. The work gets better. Then the metered bill arrives at scale, and the conversation flips from "how do we do more of this" to "how do we do less of this, somewhere cheaper."
So the teams that were just enabled start looking for the exit. They train their people on Gemini and Amazon's tools and begin shifting work there, largely because those remain subsidized and the cost is easier to absorb. Whoever subsidizes most is buying market share right now, and Google, Amazon, and the Chinese labs all appear content with that trade. When consumption pricing prices people out of the tool they were just trained on, the company selling it is throttling its own adoption at the worst possible moment.
The constraint is predictability
The easy read is that companies are balking at the cost. That isn't what we see. Companies are not refusing to spend money on AI. They're refusing to spend money they can't predict.
The evidence sits at both ends of the market. On the low end, a company will pilot Claude on something like a thirty-dollar monthly budget per person, which a genuine user exhausts in an afternoon. On the high end, the same company will happily hand an AI engineer a real tools budget, yet nobody can say going in what the meter will read at month's end. So the guidance that filters down to the team becomes some version of "go easy on it, we don't know what this costs." That instruction is the bottleneck, and it has nothing to do with the size of the number. It's the uncertainty. Every enterprise already knows how to buy software by the seat. They take a per-seat price, multiply by headcount, and put it in a budget without a meeting. Consumption pricing removes that certainty and replaces it with a figure that only appears after the money is spent, so the safe move becomes using the tool less.
Make the model cheap, make the harness expensive
The fix follows directly. Make the raw model cheap enough that using it is a non-decision, and put the price on the harness, sold by the seat, at a level a company can predict and defend.
Cheap model access does more than lower a bill. It keeps builders plugged into a frontier model instead of leaving for it. If building on a lab's model costs almost nothing, developers build on it. If model access is expensive, they start weighing open-source weights they host themselves or a stack they assemble by hand, which is slower, weaker, or in some cases genuinely risky. Cheap access is how a lab keeps developers building on its own model instead of defecting to whatever is free, including the increasingly capable open-weight models coming out of China.
Then charge real money for the harness, because the harness is the value. The model is close to a commodity now. The harness, the memory, the tools, the projects, the workflow that turns a raw model into something a team depends on daily, is the part that is hard to copy. Anthropic's is, in our experience, the best of them. Today's pricing has this backwards. It treats the harness as free and lets the commodity underneath run up the bill. The current structure charges for the commodity and gives away the moat, which is a strange way to defend a lead.
The labs already have the shape of the alternative. Their consumer tiers began near $20, $100, and $200 a month. Carry that seat-based logic into the enterprise and raise it by roughly an order of magnitude, and you land on something like $250, $1,000, and $2,000 a seat: a base tier for most knowledge workers, a power-user tier, and a builder tier for the engineers who create the workflows everyone else runs on.
The multiple isn't invented
That order-of-magnitude step is not a number pulled from the air. For years the ceiling on mainstream software sat around $150 to $175 per user each month, the price of Salesforce's Enterprise edition, the system a company's business ran on. Its Unlimited tier is now $350, and its new AI-bundled tiers, Agentforce and Einstein, list at $500 to $550 a seat. The ceiling has already moved, and it is moving because of AI.
Against that backdrop the harness numbers are less exotic than they first sound. Put an entire knowledge workforce on the $250 base seat and the cost lands near $3,000 per person a year. Benchmarks put total software spend at roughly $9,000 to $14,000 per employee, so even at full scale the base harness runs between a fifth and a third of what a company already spends per head on software, and some of it replaces line items already on that list. A $2,000 seat is a larger commitment, but it buys a person who can stand up a good-enough version of software a company used to purchase outright. We don't yet see enterprises writing checks of this shape for AI, and we're not claiming they do. The point is that the structure sits in the same order of magnitude as the budget already beside it, and unlike a meter, it's a figure a CFO can plan around in January and not be surprised by in December.
The strategy isn't wrong. The timing is.
To be fair to the labs, OpenAI and Anthropic are moving toward public markets and need to show a credible path to profitability, and consumption pricing is one way to show it. On its own terms that is a defensible choice. The problem is the moment. There is still a land grab underway, daily use is spreading quickly, and the companies that lock in that daily habit now are the ones who keep it later. Metering usage down to prove margins ahead of an offering hands that habit to whoever is still willing to subsidize it.
The two leaders are not equally exposed. Our read is that Anthropic's product is valuable enough to weather a difficult pricing stretch and come out intact. OpenAI looks more exposed, because we don't see it delivering the same value for the same spend. Gemini and Amazon look strong today largely because they are still subsidizing, and the same question lands on them the day they stop.
This is a view from where we sit, across the companies we work with, and it may look different from inside a lab or a year from now when the subsidies dry up. But it is clear enough to state plainly. The model is the commodity. The harness is the value. Price them like it.
Sources
- Salesforce Sales Cloud edition pricing (Starter $25, Pro $100, Enterprise $175, Unlimited $350): Salesforce Sales Pricing
- Salesforce AI-bundled tier pricing (Agentforce and Einstein, $500 to $550 per seat): Salesforce Agentforce Pricing
- Salesforce 2025 across-the-board price increase: Salesforce pricing update
- Enterprise software spend of roughly $9,000 to $14,000 per employee per year: ITBudgetCalculator, IT spend per employee and Cledara, average SaaS spend per employee 2026
Related reading
I've been circling this from a few angles. If it resonated, two earlier pieces go deeper on the two halves of the argument:
- The Thing That Holds the Thing, on why the harness, not the model, is where the value and the defensibility actually live.
- I Watched a Team Talk About Their Work in Dollars, on what consumption pricing does to the people at the keyboard, and why the meter quietly suppresses the exploration that creates the value.