OpenAI's Pricing Tells Two Stories at Once

In early August, OpenAI CFO Sarah Friar made the industry's new case.

A more capable model may have more expensive tokens, but complete the same task in one pass. — Sarah Friar

The argument is elegant: don't look at the per-token price. Look at the per-task cost. A premium model that gets it right the first time is cheaper than a cheap model that needs five tries. Days later, OpenAI slashed the fees on its Luna model by 80% ^[1]. The two facts do not fit together. If premium models were genuinely cheaper per completed task, customers would not need an 80% discount to keep using them. The price cut is not a refinement of the cost-per-task logic. It is a refutation of it, delivered by the same company in the same week. The evidence that customers are not buying the argument is already in. Corporate card data from Ramp shows users migrating to cheaper and open-source alternatives.

A lower-cost model may have cheaper tokens, but getting great results may require more attempts, more time, or more human review. — Sarah Friar

Anthropic, which joined OpenAI in pushing the cost-per-task metric ^[2], introduced its own high-performance model at half the price of its top system in the same round of cuts ^[1]. The metric push and the price war are not separate strategies. They are two responses to the same pressure, and they point in opposite directions. The cost-per-task argument is one of four defensive maneuvers the leading AI labs have undertaken in recent months. Read in sequence, each concedes more ground than the last. The first was the price cuts themselves — an acknowledgment that the sticker price on premium models had become an obstacle to adoption, whatever the per-task math might say. The second was the retreat from consumer to enterprise. In March, OpenAI declared "code red" and pivoted away from a "do everything" strategy, shuttering its Sora video product, its Atlas browser, and its e-commerce features ^[3]. Fidji Simo, who led the restructuring, was blunt about the reasoning.

We cannot miss this moment because we are distracted by side quests. — Fidji Simo

The pivot followed a period in which OpenAI abandoned consumer-facing products one after another, while Apple and Google moved in the opposite direction — pursuing mass consumer AI ^[4]. The labs that built the technology are no longer trying to sell it to everyone. The third maneuver is the turn toward restricted access. OpenAI and Anthropic are expected to implement premium pricing tiers, identity verification requirements, and government licensing frameworks — creating artificial scarcity where the technology itself offers none ^[5]. The logic is the same logic behind the enterprise retreat: if you cannot make the economics work at scale, shrink the scale. The fourth is the cost-per-task metric itself — an attempt to change the question rather than answer it. When the per-token price is falling and total token consumption is exploding, reframing value around task completion is a way to stop the conversation about unit costs before it reaches the balance sheet. None of these maneuvers addresses the condition they are all responding to. That condition is visible in two numbers. The first is 95%. MIT and London Business School researchers found that 95% of enterprise generative AI pilots fail to produce measurable P&L impact ^[6]. The failure is not a failure of the technology at the task level. It is a failure at the enterprise level — the gap between a worker completing more tasks and a company making more money. The second is 27%. Anthropic's own study of its Claude Code tool found that AI enabled employees to complete tasks that 27% of staff would otherwise have skipped ^[7]. The technology works at the task level. The MIT study says it does not work at the P&L level. Both findings can be true at once, and the space between them is where the industry's economics are breaking down. That breakdown is not contained within the AI labs. It is radiating outward into the software industry they are supposed to be disrupting. In February, Mistral AI CEO Arthur Mensch made a prediction that would trigger a sell-off.

more than half of what's currently being bought by IT in terms of SaaS is going to shift to AI — Arthur Mensch

The prediction cut ServiceNow in half from its highs, with Toast down 44% and DoorDash down 38% ^[8]. Anthropic's April product launches — Claude Managed Agents and Cowork plugins — triggered a second wave, hitting ServiceNow, Salesforce, Qualys, and Zscaler simultaneously ^[9]. The disruption is real enough to destroy value. What it has not yet demonstrated is that it can create it. The SaaS companies being disrupted are carrying roughly $150 billion in leveraged debt, much of it maturing between 2026 and 2029 ^[10]. Creditor takeovers have already begun: Blackstone seized Medallia from Thoma Bravo after four years of deferred interest on $2.8 billion; Vista Equity wrote off its investment in Pluralsight ^[11]. The firms whose revenue AI is supposed to absorb are already failing to service their debts. The firms doing the absorbing, meanwhile, have not found a way to make the absorption pay. As prior reporting on the labs' finances has shown, the two flagship AI labs lose roughly 59 cents on the dollar. Goldman Sachs forecasts that monthly token consumption will increase 24-fold to 120 quadrillion by 2030, and the cost burden is already visible: Uber exhausted its annual AI coding budget within months, and Microsoft restricted engineers from using third-party coding tools ^[12]. Simon Gooch of Saviynt has stated that long-term cost models for AI currently lack a logical basis ^[12]. The unit economics that would make this disruption sustainable do not yet exist. The four maneuvers — price cuts, enterprise retreat, restricted access, and metric redefinition — are not a strategy. A strategy assumes a path to a sustainable outcome. What the labs are doing is discovery: finding out, in real time, that the unit economics to make this disruption sustainable do not yet exist.