← Back

The Return on Tokens

16 min read · July 2026

Why AI access will broaden—and AI advantage will not

Intelligence Now Has Two Prices

One of the most revealing charts in artificial intelligence right now is not a traditional model leaderboard.

It is a cost curve.

The chart plots performance on DeepSWE, a benchmark of original, long-horizon software-engineering tasks, against the average dollar cost of attempting each task. The ideal position is the upper-right corner: higher performance at lower cost.

But the points do not form a clean hierarchy.

Some expensive models perform well. Some much cheaper models approach them. Some models consume more tokens, run longer, and cost more without solving more tasks. Across the agents tested, output tokens, runtime, and dollar cost varied by roughly an order of magnitude, yet none of those measures strongly correlated with success.

DeepSWE is a coding benchmark, not a universal ranking of intelligence. Its results should not be treated as a verdict on every model, product, or kind of knowledge work.

But the economic signal is important.

The price of a model is becoming less informative about the value of the outcome it produces.

That is happening at the same time as another, apparently contradictory trend.

Artificial intelligence is becoming dramatically more expensive to create.

One analysis estimated that the amortized cost of training the most compute-intensive models had increased approximately 2.4 times per year since 2016. Extending the frontier increasingly favors a small number of companies with access to enormous amounts of capital, infrastructure, data, and talent. The underlying research estimates both hardware and labor costs, rather than treating compute as the whole expense.

Yet the cost of purchasing an existing level of capability is moving in the opposite direction.

A recent analysis found that the price of achieving a fixed level of performance across knowledge, reasoning, mathematics, and software-engineering benchmarks had fallen roughly fivefold to tenfold per year. At the same time, the cost of running the most capable frontier systems can rise as larger models and greater reasoning effort are used to obtain marginal improvements at the edge.

That is the paradox at the center of the emerging token economy.

Artificial intelligence is becoming more expensive to invent and cheaper to use.

The production of frontier intelligence is concentrating.

The consumption of capable intelligence is commoditizing.

And competitive advantage is beginning to move into the space between them.

Commoditization Happens One Frontier at a Time

To say that models are becoming commoditized does not mean that all models are becoming identical.

The frontier still matters.

The last few points of performance can be enormously valuable in scientific research, difficult engineering, high-stakes analysis, or any other task where a small improvement in reliability creates a large improvement in economic value.

Those final points may also be extremely expensive.

But commoditization does not happen to all intelligence at once.

It happens one capability frontier at a time.

A capability first appears inside a small number of expensive systems. Competition spreads it. Distillation compresses it. Hardware improves. Open models reproduce more of it. Smaller models absorb it. What was previously available only at the frontier becomes available through many providers, at lower prices, in smaller systems, and eventually on local devices.

Yesterday’s breakthrough becomes today’s feature.

Today’s premium becomes tomorrow’s baseline.

This creates a barbell market for intelligence.

At one end, a small number of laboratories spend extraordinary amounts of capital to extend the frontier.

At the other, a growing number of providers compete to deliver capabilities that have already diffused.

The middle becomes increasingly crowded.

This does not eliminate model advantage.

It relocates it.

When access to a capable model is rare, the model itself can be the moat.

When many people can access models with similar capabilities, the model becomes an input into a larger production system. The advantage shifts toward the way that input is selected, directed, combined, verified, and converted into action.

Cheap electricity did not make every factory equally productive.

It increased the importance of factory design.

Abundant machine intelligence will do something similar.

As intelligence becomes easier to purchase, the systems surrounding intelligence become more important.

A Token Is Not a Unit of Value

The AI industry sells tokens because tokens are measurable.

Businesses do not ultimately buy tokens.

They buy resolved customer issues, working software, completed analyses, better decisions, faster research, lower risk, and additional revenue.

The token is the billing unit.

It is not the economic unit.

A million tokens from one model do not necessarily contain the same useful intelligence as a million tokens from another. Models use different amounts of internal reasoning. They require different numbers of retries. They fail in different ways. They impose different levels of review. They may produce outputs of similar apparent quality with radically different reliability.

Even the listed price can be misleading.

A recent study of reasoning-model pricing found cases in which the model with the lower advertised price produced a higher total inference cost, driven by differences in reasoning-token consumption. It is a preprint, so its exact results should not be generalized to every workload, but it exposes the weakness of comparing systems through price lists alone.

The more useful concept is not cost per token.

It is cost per successful outcome.

Researchers have begun describing this as the cost of pass: the expected monetary cost of generating a correct solution. Their results show that the most economical model depends on the task. Lightweight models can be most efficient for basic work, larger models for knowledge-intensive tasks, and expensive reasoning systems for difficult quantitative problems. More intelligence is not always more economical, but cheaper intelligence is not always cheaper either. Cost-of-Pass provides a useful framework for thinking about this tradeoff.

The business metric should go one step further: return on tokens equals verified economic value created, divided by the complete cost of inference, tools, review, rework, latency, and expected errors.

This denominator matters.

A cheap answer that requires an hour of expert verification may be expensive.

An inexpensive model that fails silently may be ruinous.

A premium model that prevents a costly mistake may be extraordinarily cheap.

And an answer that is correct but never enters a decision, workflow, or product has almost no economic value at all.

The price of intelligence is not what a provider charges to generate the output.

It is what the organization spends to obtain, trust, and act on the outcome.

That is why token volume will become a poor measure of AI adoption.

A company consuming more tokens may be more productive.

It may also be producing more drafts, more retries, more unverified outputs, and more automated waste.

Tokens measure activity.

They do not measure consequence.

Access Is Becoming the Wrong Frame

The first phase of enterprise AI has been organized around access.

Which employees have a license?

Which model can the company use?

Which vendor has the strongest benchmark scores?

How many tokens should each team receive?

Those questions were reasonable when capable intelligence was scarce.

They become less useful as model capability diffuses.

Two people can now use the same model, with the same subscription and the same token budget, and produce radically different economic results.

One person asks for a generic answer.

Another identifies a valuable problem, assembles the relevant evidence, supplies proprietary context, constrains the output, tests its assumptions, recognizes uncertainty, and integrates the result into a consequential decision.

The model is the same.

The return is not.

That difference is conversion capability: the ability to convert machine intelligence into verified economic value.

It begins with problem selection.

A powerful model applied to a trivial problem still creates trivial value. A modest model applied to a frequent, expensive, and well-specified bottleneck may create enormous value.

It continues with context.

Models do not know which internal data is authoritative, which customer matters, which exception applies, which constraint is political rather than technical, or which apparent optimization would damage another part of the business. Intelligence without context can be impressive and directionless at the same time.

Then comes allocation.

Not every problem deserves the strongest model. Not every task deserves a long reasoning trace. Not every ambiguity deserves ten attempts. The economically capable user assigns the right level of intelligence to the right problem.

Research on model cascades has already demonstrated the potential of this allocation layer. In benchmark experiments using the models available at the time, FrugalGPT matched the performance of the strongest individual model while reducing costs by as much as 98% in some settings. The exact savings will not carry across every production environment, but the mechanism is durable: intelligent routing can be more valuable than indiscriminately purchasing the strongest model.

Then comes verification.

The user must know what evidence would establish that the answer is correct, what failure would look like, and when the model has crossed from useful inference into plausible fabrication.

Finally, there is execution.

A recommendation without authority, integration, distribution, or accountability remains text in a window.

That produces a distinction between nominal access and economic access.

Nominal access means that a person can call the model.

Economic access means that the person can use it to produce a verified outcome whose expected value exceeds its complete cost.

Nominal access will broaden.

Economic access will remain uneven.

AI Compresses Execution Before It Rewards Judgment

There is an important counterargument to the idea that existing capability simply multiplies AI advantage.

In several workplace studies, the largest gains from generative AI have gone to less experienced workers.

A study of 5,179 customer-support agents found that AI assistance increased productivity by 14% on average, including a 34% improvement among novice and lower-skilled workers, while having little effect on the most experienced workers. The researchers found suggestive evidence that the system was distributing the practices of stronger workers and helping newer employees move down the experience curve more quickly.

That result does not weaken the return-on-tokens thesis.

It makes the thesis more precise.

AI can compress the execution gap.

It can take practices that were previously stored in the experience of a small number of people and distribute them across a much larger workforce. It can help a novice produce a competent response, a reasonable analysis, or a structured first draft much sooner.

This makes some forms of expertise less scarce.

But it does not make judgment disappear.

In an experiment involving 758 consultants, AI users completed more tasks, worked faster, and produced substantially higher-quality outputs when the assignments fell inside the model’s capability frontier. On a task deliberately placed outside that frontier, however, consultants using AI were materially less likely to reach the correct answer.

The distinction is crucial.

There is the capability to perform the work.

And there is the capability to govern the work.

Governing the work means knowing which problems are appropriate for AI, what information the model is missing, which answer requires verification, which uncertainty matters, and when the human should take control.

AI can make the production of an answer abundant while making the selection and evaluation of answers more valuable.

It can commoditize execution while shifting scarcity toward judgment.

That is why the most important future skill may not be traditional expertise alone, or “prompt engineering” as a separate technical craft.

It may be the combination of domain knowledge and AI judgment: enough expertise to frame the problem, recognize the boundary of the system, inspect the result, and accept responsibility for what happens next.

As the cost of producing an answer falls, the value of asking the right question rises.

As drafting becomes abundant, taste becomes more important.

As analysis becomes cheap, decision quality becomes the bottleneck.

AI will commoditize more answers.

It will premiumize better questions.

This also creates an apprenticeship problem.

Most professionals did not develop judgment by studying judgment in the abstract. They developed it by doing the lower-level work: drafting, calculating, researching, debugging, comparing, and being corrected.

If AI removes too much of that work, organizations may improve today’s output while weakening the process that creates tomorrow’s experts.

AI can compress the learning curve.

It can also remove it.

The organizations that benefit most will not merely automate junior work. They will redesign apprenticeship so that people still practice prediction, diagnosis, review, and explanation even when the machine can produce the first answer.

The capability multiplier has to be built somewhere.

The Organization Is Part of the Model

Individual capability is only part of the equation.

A skilled employee inside a poorly designed organization can still generate very little value from AI.

The employee may lack access to the relevant data.

The model may be prohibited from using the necessary tools.

The output may require approval from five people.

No one may know who is accountable for verifying it.

The recommendation may arrive faster, only to wait inside the same slow decision process.

Research on earlier general-purpose technologies suggests that this is normal. Their full productivity effects usually require complementary investments in skills, workflows, organizational structures, and business models. Those investments often create an implementation lag between the appearance of a powerful technology and its visible economic return. This productivity-paradox analysis explains why those complements matter.

A six-month randomized experiment involving 6,000 workers illustrates the boundary. Employees who actively used an AI tool integrated into their existing applications spent approximately three fewer hours per week on email. But their time in meetings did not significantly change. AI could alter a task the individual controlled independently. It could not, by itself, redesign a coordination process shared with the rest of the organization. The study is a useful reminder that task acceleration is not organizational redesign.

The model can compress a task.

It cannot automatically rewrite the organization around the task.

That is why the organization itself becomes part of the intelligence system.

The organization determines what context the model receives.

It determines which tools the model can use.

It determines when the model should escalate to a person.

It determines what evidence counts as verification.

It determines who can act on the output.

And it determines whether the result of one task is retained as learning for the next.

Two companies can license the same model and obtain completely different economics because one has built the surrounding system and the other has merely purchased access.

The strongest AI organization may not be the one with the best individual prompters.

It may be the one that turns expert judgment into reusable infrastructure: routing policies, context systems, evaluation suites, approval rules, memory, feedback, and clear decision rights.

In Recursive Enterprise Intelligence, I argued that the strategic control point is moving away from model access and toward the improvement loop surrounding the model: the traces, evaluations, context, adaptations, and workflow-local learning generated through use.

The return-on-tokens thesis is the economic counterpart to that argument.

The improvement loop matters because it changes the return on every future unit of intelligence.

A correction can become context.

A failure can become an evaluation.

An override can become a routing rule.

A repeated exception can become a new requirement.

The same workflow that produces an outcome can also produce the evidence required to improve the next outcome.

That is the difference between consuming intelligence and compounding it.

Productive Access Can Compound

Once an individual or organization can demonstrate a high return from AI, another effect becomes possible.

Investment begins to follow the return.

A team that repeatedly converts AI into measurable revenue, lower costs, faster execution, or better risk control can justify better tools, deeper integrations, larger reasoning budgets, more proprietary context, and greater autonomy.

A team that cannot demonstrate value remains stuck at the level of generic assistants and isolated experiments.

This creates a potential flywheel:

Better problem selection produces more valuable outcomes.

More valuable outcomes justify greater investment.

Greater investment creates better context, tooling, evaluation, and workflow design.

Those systems then increase the return on the next use.

The process is not automatic.

Additional compute can have diminishing returns. Organizations can mistake activity for value. A larger budget can simply fund more sophisticated waste.

But where returns can be measured, machine intelligence begins to behave like other forms of productive capital.

It flows toward the people and systems that appear able to use it well.

This means that falling model prices may not reduce total AI spending.

They may expand it.

When a unit of intelligence becomes cheaper, companies can apply it to more tasks, run it for longer, explore more alternatives, and delegate larger portions of workflows. The cost of a fixed capability can fall even while the scale of machine work rises.

Unit intelligence gets cheaper.

Cognitive workloads get larger.

The most capable users will not necessarily be those who spend the most on a single model.

They will be those who know where another dollar of intelligence is likely to create more than another dollar of value.

That is a capital-allocation problem.

It is also a capability problem.

The next divide may therefore be self-reinforcing.

People and organizations that produce high returns from models can invest in systems that raise those returns further. Those that treat AI as an undirected stream of inexpensive content may receive the same nominal access without developing the same productive access.

Everyone may be able to call a model.

Not everyone will be able to justify the next token.

The Return on Tokens

The token economy will reward a different set of behaviors from the software economy that preceded it.

For individuals, the advantage will not come from producing the largest quantity of AI-assisted output. Output itself is becoming abundant.

The advantage will come from identifying important problems, constructing the right context, recognizing weak reasoning, making consequential decisions, and taking responsibility for results.

For companies, the central metric should not be the number of licenses distributed or tokens consumed.

It should be verified value per total cost.

Which workflows improved?

Which decisions became faster or better?

Where did the model fail?

How much human review was required?

What happened after the output was produced?

And did the organization learn anything that will make the next run better?

For model providers, the market will continue to contain both premiums and commodities.

The frontier can remain scarce and expensive even as capabilities behind it become widely available and increasingly substitutable. The strongest labs may continue earning substantial returns from extending the edge. But every capability that diffuses will move part of the competitive battle away from the model and into the systems that allocate and apply it.

For educators and managers, the objective cannot be merely to teach people how to obtain answers from AI.

It must be to teach them how to evaluate answers, how to locate the frontier of the system, how to work with uncertainty, and how to preserve the development of judgment when the machine can perform more of the execution.

Most companies will initially treat the token economy as a procurement problem.

Which vendor?

Which model?

Which price?

Which contract?

But the deeper question is not which intelligence the company can purchase.

It is what the company can turn that intelligence into.

Artificial intelligence is becoming more expensive to create and cheaper to use.

That will not make advantage disappear.

It will move advantage from owning the best model to earning the highest return from models.

The winners will choose more valuable problems.

They will provide better context.

They will allocate the appropriate level of intelligence.

They will verify what the system produces.

They will act on the result.

And they will make every use improve the economics of the next.

Nominal access to intelligence will broaden.

Productive access will not broaden automatically.

Models are becoming abundant.

The ability to turn them into verified, compounding value remains scarce.

The next AI divide will not be access to intelligence.

It will be the ability to convert intelligence into consequence.

Sources