Sam Altman almost scrapped token pricing. Here’s why he backed off.

Axiom considered abandoning token-based pricing for its next model. Instead, a developer would submit a prompt and receive a price for the completed work. Sam Altman compared the proposed system with AWS, where usage happens first and a bill arrives later. The change mattered because it could have simplified billing while making costs less visible.

Axiom tested customer reactions and found strong resistance. Buyers said they needed to know what they were paying for. Developers already track input and output tokens to control spending and estimate future usage. A single unexplained charge would make that planning harder, especially for workloads with changing demands.

Altman said customers would even accept higher costs if pricing stayed tied to consistent tokens. They trusted the visible unit more than an opaque job price. For now, Axiom continues using token pricing. The company acknowledges tokens are imperfect but has not presented a replacement that customers prefer.

Token-based pricing meters an AI exchange through tokens, the text units used in a prompt and the model’s response. Developers are charged according to input and output usage. This gives them a visible way to connect a bill with what the model processed and produced, even though Sam Altman calls tokens a poor measurement of AI work.

A flat-fee system would work differently. A developer could submit a prompt, and Axiom would return a price for the work rather than charge token by token. Altman compared that arrangement with AWS: use the service, then receive a bill. The charge would cover the task without necessarily revealing how Axiom calculated it.

That distinction matters for budgeting. Token pricing lets developers monitor usage, compare jobs, and estimate costs as activity grows. A task price could be simpler, but customers feared it would hide the mechanism behind each charge. Their negative response led Axiom to back away.

Developers want input and output token counts because those figures show what generated a bill. Input tokens represent the material sent to a model, while output tokens reflect the response produced. Tracking both helps developers see how much their applications consume and identify where spending is growing.

The article says developers already spend substantial time monitoring these amounts. They use token usage to keep costs down and estimate future expenses as usage increases. For example, a longer prompt or a larger response can change the charge. Seeing the components lets a team connect that change with actual activity rather than guess.

A single opaque price would remove that visibility. Customers told Axiom, “I need to know what I’m paying for somehow.” They preferred paying more under consistent token pricing to receiving an unexplained charge. This transparency is especially important for agents, whose multistep workflows can consume unpredictable amounts of compute.

Sam Altman says the cost of obtaining a given level of AI falls roughly tenfold every year. In practical terms, the same level of capability becomes far cheaper to provide over time. The article presents this claim as part of a broader discussion about whether token pricing remains the right way to measure AI usage.

That cost decline does not solve the billing problem. Developers still need to understand what each model call costs and how usage changes over time. They may also face different charges when choosing between a cheaper model and a more capable one. Guessing how much reasoning a task demands adds further cognitive overhead.

Axiom therefore remains caught between a changing cost base and an imperfect pricing unit. The company considered charging by completed task, but customers rejected opaque bills. For now, token pricing remains in place, even as Altman acknowledges its shortcomings and says AI costs are falling rapidly.

Developers value token pricing because it gives them a visible unit behind the bill. Even if tokens are imperfect, they can count inputs and outputs, compare usage, and connect spending with activity. That visibility creates a sense that charges reflect something measurable rather than an unexplained decision by the provider.

Altman said people he consulted would “much rather” see tokens inflated and pay more for the same output. Their reason was simple: “I know there’s a real thing there.” By contrast, a single job price would not show how the amount was calculated. Developers might struggle to compare one task with another or understand why costs changed.

Altman believes this distrust drives the preference for transparent units. Developers suspect they could “get screwed somehow” over the long run under an opaque system. The reaction persuaded Axiom to keep token-based pricing, despite its acknowledged flaws and added cognitive overhead.

AI agents are applications that carry out tasks through multiple connected steps. Instead of handling one simple exchange, they may repeatedly use a model as a workflow progresses. The article focuses on developers building agents because their workloads can consume varying amounts of compute before a job is complete.

That variability makes budgeting difficult. One agent task may require more steps, more model calls, or more reasoning than another. Each part can add usage, so the final resource demand is not always clear at the start. Even token-based pricing already makes these workflows difficult to forecast because the amount of work can change during execution.

A flat job price might appear attractive for agents because it could hide that complexity. However, Axiom did not explain how it would set those prices or whether customers could compare one charge with another. Developers’ rejection of opaque billing means agents currently remain under token-based pricing, with unpredictable costs still a concern.

Tokens are imperfect because they count pieces of an exchange rather than the full work required to produce an answer. Sam Altman called them “a kind of fake unit”: real in one sense, but not a good unit. Token totals do not fully show how much processing, model capability, or reasoning a task demands.

The underlying resource is compute. A request can involve different amounts of computational work even when its token count does not clearly reveal that difference. The article also notes that developers must guess how much reasoning a task will require. Choosing between a cheaper model and a more capable one creates additional cognitive overhead without necessarily adding value.

Axiom has not found a better measure that customers trust. Altman rejected the idea of an intelligence-equivalent token and compared today’s token debate with chipmakers’ nanometer marketing. For now, tokens remain the billing unit, even though they only approximate the compute behind AI output.