The cost of using AI models is declining, even as the number of requests continues to grow rapidly. Companies are shifting to cheaper models, OpenAI and Anthropic are cutting prices, and AI agents are increasingly using caching mechanisms that reduce costs. This is good news for businesses, but for an industry investing hundreds of billions of dollars in computing infrastructure, it raises an uncomfortable question: how sustainable will the AI boom be if consumption of computing resources grows faster than the revenue generated from their use?

Tokens are becoming almost twice as cheap

The cost of tokens — units of text and other data processed by AI models — has dropped noticeably in recent months. Several public indexes show a similar trend, although there is no single metric that reflects the entire market.

The Silicon Data token spending index, which takes into account the usage volume of different models and the changing balance between them, recently fell below $1 per million tokens. At the end of May, the indicator reached $2.07, meaning it has declined by 54% from its peak.

The index from Ramp, a company specializing in corporate expense management and tracking actual business spending on OpenAI and Anthropic products, shows a similar picture. Compared with the peak at the end of February, when one million tokens cost an average of $1.22, the figure has fallen by 48%.

Several processes are driving these numbers at once. Companies are increasingly routing requests to cheaper models, including open-weight models, while AI developers themselves are lowering prices as new generations of products are released.

Companies no longer want to pay for the most powerful models

Previously, one of the symbols of the AI boom was companies’ willingness to spend ever more money on access to the most powerful models. Now businesses are learning to allocate requests more efficiently: not every task needs to be assigned to the most expensive system.

According to Steve Howe, head of research at Silicon Data, part of demand is shifting from cutting-edge models to more affordable open alternatives. This means that growth in AI usage no longer necessarily leads to a proportional increase in spending on the most expensive services.

Competition among providers is also putting additional pressure on prices. New model versions offer more capabilities at lower cost, so developers have to cut prices to retain customers.

At the same time, corporate AI spending can quickly shift between companies. Ara Kharazian, Ramp’s chief economist, notes that such spending still cannot be considered firmly locked in with any specific provider: the release of a new model can redirect billions of dollars in spending from one service to another.

For developers, this means that even a large customer base does not guarantee preservation of previous revenue. If a competitor offers comparable quality at a lower price, customers may change their preferences fairly quickly.

AI agents are also cutting costs

Another reason for the drop in prices is the spread of AI agents, which conduct extended dialogues, perform chains of actions, and repeatedly call on models within a single task.

As Chris Clark, co-founder and COO of the OpenRouter platform, explains, a significant portion of the context in such interactions can be stored in cache rather than processed again with each request. Tokens corresponding to cached input data are usually significantly cheaper than standard ones.

As a result, an agent can continue working with a large volume of context without paying again for all of its processing at the standard rate. This makes extended interactions more economical, but at the same time reduces the average cost of the tokens being used.

Thus, growth in the number of requests and longer AI operation do not necessarily translate into comparable growth in revenue for model providers.

Trillion-dollar investments amid stagnant spending

For the tech industry, falling prices are not necessarily a bad sign in themselves. More affordable AI can stimulate the emergence of new applications, broaden the user base, and increase overall demand. This effect is known as Jevons paradox: improving the efficiency of resource use can lead to higher overall consumption of that resource.

So far, this logic is working in at least one respect: the volume of AI usage, measured by the number of tokens, continues to rise rapidly.

But Ramp’s data reveals a more troubling divergence. In August and September, corporate spending on tokens barely grew for two consecutive months, even though actual AI usage continued to increase.

“For the first time, we are seeing token usage grow while spending on them remains at the same level,” Kharazian says. According to him, this situation has now persisted for two months in a row.

This distinction is fundamental. Growth in the number of processed tokens shows that AI is being used more and more actively, but it does not answer the question of how much money the industry is earning from that usage. If businesses are solving more tasks for the same money, model providers have to look for other ways to increase revenue.

Meanwhile, the scale of infrastructure investment remains enormous. According to Goldman Sachs, the largest U.S. hyperscalers — operators of cloud platforms and large-scale data centers — will direct about $800 billion in capital expenditures in 2026. In 2027, their combined investment is expected to reach $1.1 trillion.

Such investments assume that demand for computing power will not only grow, but also provide sufficient economic returns. If AI usage continues to rise without comparable growth in customer spending, it will become harder to justify the scale of infrastructure investment.

What this means for Anthropic and the AI market

Declining token costs do not yet prove that demand for AI has hit a ceiling or that the investment boom is nearing its end. It may reflect improved model efficiency, intensified competition, and customers shifting to more cost-effective tools. If lower prices lead to even broader adoption of AI, the overall market may continue to grow over time.

However, for investors it is important to watch not only the number of requests and computations, but also how much companies are willing to pay for them. Corporate spending is what is closely tied to the revenue of model developers — and therefore to their ability to justify high valuations ahead of a potential IPO.

For companies like Anthropic, this is especially important amid expectations of major IPOs in the AI sector. If customer spending begins to consistently lag behind usage volumes, the market will have to reconsider assumptions about how quickly model developers can turn growing demand into revenue.

The main question now is not only how actively AI is developing, but also whether the industry will manage to earn more from technologies that are becoming cheaper with each passing month.