The price of tokens is falling while their consumption is growing rapidly. Behind this new artificial intelligence economy lies a much more strategic question for companies: when intelligence becomes abundant, where is value still created?
Artificial intelligence is getting cheaper, much cheaper even. According to data published by Ramp and reported on October 28, 2026 by Axios, the weighted average price paid for OpenAI and Anthropic tokens reached approximately $0.59 per million tokens on October 23, compared to $0.77 four weeks earlier. Compared to the end of June, the decrease is around 35%. At first glance, the conclusion seems simple: companies will be able to use more artificial intelligence for less money. That’s correct, but it’s probably not the most important economic consequence. Because at the same time, companies are consuming much more AI. Ramp observes a 37.1% increase in token volume over four weeks, while associated spending only increases by 8.9%. We are therefore facing a paradox that could profoundly transform the AI economy: the unit cost of intelligence is decreasing, but its consumption is increasing rapidly. For executives, the question is therefore no longer just how much artificial intelligence costs, but what happens when it becomes cheap enough to be used everywhere.
A cheaper resource doesn’t necessarily mean lower spending
Economic history shows that a dramatic improvement in a technology’s efficiency doesn’t always lead to reduced consumption. It can produce exactly the opposite effect. When a resource becomes cheaper, new uses appear, what previously seemed too expensive becomes possible, then commonplace. This is potentially what is happening with artificial intelligence. A company that yesterday used a model to summarize a few documents can tomorrow analyze its entire internal documentation. A customer service can go from a few AI-assisted responses to thousands of automated interactions. A development team can multiply code generations, tests, simulations and controls. The price drop therefore doesn’t necessarily mean the company will save its budget. It mainly means it can buy much more capacity for a comparable amount. Cheaper AI doesn’t necessarily mean less spending; it means more possibilities.
Agents completely change the equation
This question becomes even more important with the emergence of agentic systems. A traditional chatbot responds to a request, while an agent can perform a succession of operations. It can search for information, consult a document database, use a reasoning model, compare multiple answers, check a result, call another tool, execute an action and then verify it was performed correctly. Each step can generate new calls to the models. The cost of a query can therefore decrease even as the cost of the complete process increases. Gartner has identified this dynamic under the term “Inference Paradox.” The improvement in the unit economics of models can be accompanied by an increase in the total cost of workflows because they become much more complex. According to Gartner, some advanced reasoning models can consume five to thirty times more tokens than simple conversational uses. The firm also estimates that inference costs associated with certain agentic workflows could be multiplied by more than five by 2028. That’s the real economic change: the token becomes cheaper, but uses become much more demanding. And this evolution leads directly to a question that companies can no longer avoid: does their additional consumption of intelligence produce equivalent economic value?
The risk: winning on price and losing on ROI
This is probably one of the most important issues for executives in the coming years. Companies could perfectly succeed in reducing the unit cost of their AI uses while degrading their overall return on investment. Gartner emphasizes that only a minority of analyzed agentic deployments today produce a truly robust return on investment. The problem isn’t necessarily the technology; it often lies in the process to which it’s applied. Let’s take a simple case. A company automates an administrative task with a chatbot that analyzes a document and provides an answer. Tomorrow, this same company could set up an agent capable of analyzing the document, searching for additional information, consulting multiple databases, verifying response compliance, requesting a second model, and then triggering an action in business software. The cost of each elementary operation may have decreased significantly, but the process will have gained considerably in complexity. Will the company have created more value? Perhaps, but it’s not automatic. This is probably where the next frontier of AI management lies. A company can win the battle of token cost and lose the one of ROI.
Intelligence becomes cheaper, infrastructure becomes gigantic
A second paradox accompanies this price drop. While digital intelligence is gradually becoming cheaper to consume, the infrastructures needed to produce it are reaching unprecedented investment levels. According to the International Energy Agency, major data center operators should devote more than $700 billion to their investments in 2026. This growth concerns data centers, accelerators, networks, electrical equipment, cooling and energy capacities. Artificial intelligence seems immaterial, but its infrastructure absolutely is not. The decrease in model cost therefore doesn’t eliminate physical constraints; it can, on the contrary, accentuate them if it leads to an explosion in demand. The more accessible intelligence becomes, the more we need computing power to produce it, and the more this power increases, the more new bottlenecks appear: energy, electrical capacity, semiconductors, advanced memory, transformers, cooling or land availability. Scarcity doesn’t disappear. It simply changes location.
What was rare yesterday could become commonplace tomorrow
For several years, having a high-performing artificial intelligence model constituted an advantage. This may no longer be enough. The OECD observes a nearly 80% decrease in quality-adjusted prices for text-to-text models between January 2024 and April 2026. In other words, at comparable performance levels, accessing artificial intelligence costs less and less. If this trend continues, access to a high-performing model could gradually become a commodity. A company could choose between several models based on price, performance, security or the nature of the task, in the same way it chooses today between several cloud providers or different software. What was rare yesterday, access to artificial intelligence, could therefore become commonplace tomorrow. The strategic question then becomes much more interesting: what will remain scarce?
Value shifts toward what cannot be easily copied
Several resources could become decisive. The first is proprietary data. A model accessible to everyone doesn’t necessarily constitute a competitive advantage; on the other hand, the ability to feed it with unique, reliable, structured and contextualized data can create much more sustainable differentiation. The second is business expertise. A model may know the general rules of a sector, but it doesn’t necessarily know the particularities of a company, its history, its customers, its processes or the weak signals that its employees have learned to recognize over the years. The third is trust. The more systems make decisions or execute actions, the more important security, responsibility, compliance and traceability issues become. The fourth is distribution: having the best technology doesn’t guarantee having the customer relationship. And the fifth, probably the most underestimated, is organizational capacity.
The real advantage could become organizational
The OECD emphasizes that economic gains related to artificial intelligence require complementary transformations in data, skills and work organization. This observation is fundamental. A company doesn’t create value because it uses many tokens; it creates value when it improves an economic result. One may want to reduce the time needed to process a file, increase revenue, improve a decision, reduce a risk, accelerate market launch, improve customer experience, reduce a cost or allow employees to devote more time to high value-added activities. Artificial intelligence can contribute to each of these objectives, but it doesn’t guarantee them. The ability to redesign processes therefore becomes essential. Adding AI to an inefficient process doesn’t automatically transform this process into a competitive advantage. In some cases, it simply allows executing a poor organization more quickly. The question is therefore no longer just technological. It becomes managerial.
The executive’s role also changes
This new intelligence economy also requires reconsidering management responsibility. When access to AI was limited, the main question was how to adopt it. When AI becomes abundant, the problem becomes different: you must decide where to use it, for which tasks, with what level of autonomy, at what cost, with which performance indicators and above all with what expected value creation. This arbitration capacity could become one of the essential leadership skills in the age of artificial intelligence. Not all tasks necessarily deserve more AI. Not all agents will produce more value. Not all automations should be carried out simply because they become technically possible. Abundance paradoxically requires more discipline.
The next battle may no longer be about models
For a long time, competition in artificial intelligence focused on one question: who will build the most powerful model? This competition will continue, but it could gradually become less decisive for user companies. The battle could shift to data, energy, infrastructure, processes, skills, governance and distribution, and especially to organizations’ ability to transform abundant technology into measurable economic results. This is perhaps the real lesson from the figures published by Ramp. The price of a million tokens is important, but it’s ultimately just an indicator. The strategic signal is much deeper: we are gradually entering a world where digital intelligence becomes less rare. When this happens, value doesn’t disappear. It migrates to what remains difficult to reproduce: proprietary data, business knowledge, trust, infrastructure, skills and the ability to profoundly transform an organization.
For executives, the essential question is therefore probably no longer: how much does artificial intelligence cost? It becomes:
What value are we capable of creating with intelligence that is gradually becoming abundant?
Because AI that’s ten times cheaper will never automatically create ten times more value. It will only create ten times more possibilities.
Artificial intelligence creates capacity. Organizations create value.
