Why Is AI So Expensive When Token Prices Keep Falling?

"Everyone told us AI gets cheaper every year. So why did our AI bill double?"

That is the question finance leaders are now asking their technology teams, and it is a fair one. The price of using AI models has fallen faster than almost any technology price in history. Yet companies that rolled out AI tools this year are discovering that their total spend is going up, sometimes faster than their budget can absorb.

I run engineering at Chronexa, where we build AI systems for mid-sized firms and pay the AI bill for our own internal systems every month. Here is why AI is so expensive even as token prices fall, where the money actually goes, and what you can do about it without slowing your team down.

Here's what matters most:

• The unit price is falling, the volume is exploding. Token prices have dropped more than 90% since 2023, but total spending has roughly doubled since late 2025, according to the Silicon Data Token Expenditure Index.
• This is a classic pattern economists call Jevons paradox: when something gets cheaper, people use so much more of it that the total bill rises.
• Most of the waste is routing, not volume. Accenture found 54% of AI requests go to a more expensive model than the task needs.
• Very few firms can measure what they get back. Only 35% of large companies in Accenture's survey can calculate cost per outcome for their biggest AI use case.
• The fix is management, not abstinence: measure cost per result, route simple work to cheaper models, set team caps and assign an owner.

Why is AI so expensive if token prices keep falling?

First, a quick definition. A token is the unit AI providers charge by, roughly three-quarters of a word. Every time a system reads a document or writes an answer, it consumes tokens, and the provider bills per million of them.

The price per token has collapsed. Epoch AI, an independent research group, measured how fast the price of reaching a fixed level of AI performance falls, and found declines ranging from 9 times to 900 times per year depending on the task, as of its March 2025 analysis. The price to match GPT-4's performance on a set of graduate-level science questions fell about 40 times per year.

So why are bills going up? Because usage is growing much faster than prices are falling. Fortune reported in June 2026, citing the Silicon Data Token Expenditure Index, that the price of a single token has dropped more than 90% since 2023, while spending on large language models has doubled since late last year. Bain & Company analysts found that token costs halved between December 2024 and December 2025, while the number of tokens consumed rose 450% over the same period. Bain summed it up neatly: "The models get cheaper. The usage gets heavier. The bill stays stubbornly high."

Apollo's chief economist Torsten Slok called it "Jevons paradox in action." The original observation dates to 1865, when the economist William Stanley Jevons noticed that more efficient steam engines did not reduce Britain's coal consumption. They increased it, because cheaper power made new uses worthwhile. "Lower cost per interaction does not mean fewer interactions," Slok wrote. Companies simply run more AI.

Where does the extra AI spend actually come from?

When I look at an AI bill line by line, the growth almost always comes from four places.

1. More people using it. The first wave of AI use was a few curious employees with chat subscriptions. The second wave is company-wide rollouts, coding assistants for every engineer, and AI built into the tools staff already use. Usage per company jumps by an order of magnitude.

2. AI agents that take many steps. An AI agent is a system that works through a whole task instead of answering one question: it reads a document, checks a system, drafts a reply, checks its own work, tries again. Each step is another call to the model. A task a person thinks of as "one request" can be dozens of calls behind the scenes. Accenture's research found that agentic workflows already account for about 11% of enterprise token spend, and that share is growing.

3. Models that think before they answer. The newest models produce long internal reasoning before giving a final answer, and that reasoning is billed too. You get better answers, but each answer can cost several times more than it did with an older model.

4. Sending simple work to the most expensive model. This is the biggest avoidable cost. Accenture's research found that 54% of AI requests are routed to a higher tier of model than the task needs, that frontier models cost 10 to 20 times more per token than mid-tier models, and that fewer than 10% of work tasks (measured across 9,368 tasks in the US government's O*NET job database) actually need frontier capability. Classifying an email or pulling a date from an invoice does not need the smartest model on the market.

It is also worth knowing that tokens are not the whole bill. The same Accenture research put tokens at 25% or less of total AI spend. The rest is software subscriptions, cloud hosting, integration work and the people who maintain the systems.

What happened at Uber, and why it matters to smaller firms

The clearest public example this year came from Uber. Fortune reported in June 2026 that Uber's chief operating officer, Andrew Macdonald, said the company used up its entire AI budget in the first four months of the year. Uber had encouraged employees to use AI tools heavily, particularly Anthropic's Claude Code, and had even built leaderboards ranking engineers by usage. Bloomberg reported that Uber then capped monthly AI spending at $1,500 per employee.

Uber's chief technology officer, Praveen Neppalli Naga, later wrote that the company treated "efficiency as an engineering problem rather than a budget problem," quadrupling the number of employees using frontier AI tools while cost per token fell, through techniques such as reusing repeated instructions instead of paying for them every time, changing which model was used by default, and showing engineers their own usage and cost. He called it "the end of the so-called tokenmaxxing era," meaning the period when companies rewarded using as much AI as possible.

You are not Uber. But the pattern is identical at a 50-person firm, just with smaller numbers. Someone enthusiastic connects AI to a high-volume process, nobody watches the meter, and the invoice arrives at month end.

Our own lesson: the AI bill that ran out quietly

I will share one of ours, because it taught me that cost and reliability are the same problem.

We run an internal outbound system at Chronexa that researches prospects and drafts emails using an AI model, billed from prepaid credit. At one point the credit ran out. The AI provider started refusing every request. The workflow around it did not treat the refusal as a failure, so it kept logging each lead as successfully processed. For days, the dashboard said everything was fine while nothing was actually being produced.

Nobody had overspent. The opposite happened: the money stopped, and the system lied about it. The fix was not a bigger budget. It was treating "the AI provider said no" as an alert that wakes a person up, and tracking cost per finished result, not just total spend. Any firm running AI in production needs both: a meter, and an alarm on the meter.

Will AI get cheaper next year?

Per unit, almost certainly. In total, probably not.

Accenture surveyed 750 senior executives at companies with more than $1 billion in revenue in July 2026 for its report "The CIO's guide to AI tokenomics." Over the next 24 months, those executives expect token prices to fall 19% while consumption rises 78%. Without changes, that works out to roughly a 44% higher bill, or about $14.4 million for every $10 million spent today. One in three organizations had already used up its annual token budget before year end.

The most revealing figure in that survey: 95% of executives said they would spend more if token prices fell by 25% or more. Only one in twenty would bank the savings. Cheaper AI does not shrink budgets. It funds new uses.

So the planning assumption for your firm should be that your AI spend will grow, and that the job is to make sure each dollar buys something measurable, not to wait for prices to fall.

How do you control the AI bill without slowing your team down?

Here are the controls I recommend to the firms we work with, in the order I would put them in place.

1. Measure cost per result, not total spend. Total spend tells you nothing. Cost per invoice processed, per client file prepared or per call answered tells you whether the system is worth it. Accenture found only 35% of organizations can calculate cost per outcome for their largest AI use case, and 59% work from directional estimates. Start with your single biggest use.

2. Send simple work to cheaper models. Classification, data pulling and routine drafting rarely need a frontier model. Production teams told Accenture that routing work to the right model tier cut serving bills by 40% to 60%. This is usually the largest saving available.

3. Stop paying twice for the same instructions. Most AI systems send the same long set of instructions with every request. Providers now offer discounts for reusing repeated content and for processing non-urgent work in overnight batches. We wrote a guide on cutting AI processing costs by half with batch processing.

4. Show each team its own cost. Accenture found a striking link between cost visibility and measured value: organizations with no cost allocation could attribute value to about 5% of their spend, while those that charged costs back to the teams using them could attribute about 32%. When people see the meter, they use AI more deliberately.

5. Set caps before you need them. Uber's $1,500 per employee per month is one model. A cap per workflow (for example, a ceiling per document processed) is often more useful for operations teams.

6. Alarm on failures, not just on spend. As our own story shows, a system that stops working can look cheaper on the bill. Watch for sudden drops in volume as closely as sudden spikes.

7. Give the bill one owner. Accenture found 48% of organizations split AI cost accountability between IT and finance. Split ownership usually means no ownership. Pick one person who owns cost per outcome for each major use.

What about data security and vendor lock-in?

For a mid-sized firm, cost control and security usually come up in the same meeting, so it is worth addressing both.

The cheapest setup is rarely the safest one by default. Consumer AI subscriptions may use your data to improve their models unless you change the settings, and they give you little control over where data is processed. For client data, run AI through an enterprise agreement with the model provider that excludes training on your data, ideally inside your own cloud account on Microsoft Azure, Amazon Web Services or Google Cloud, with logs of what was sent and what came back.

That setup also helps with cost. When the system runs in your own account, you see the real usage bill directly, you can switch models without rebuilding everything, and no vendor can quietly mark up the AI usage they resell to you. If an AI vendor bundles usage into a flat fee, ask what happens when your volume doubles, because it probably will.

Frequently Asked Questions

Why is AI so expensive for businesses?

Because usage grows much faster than prices fall. Token prices have dropped more than 90% since 2023, yet company spending on AI models has roughly doubled since late 2025, according to the Silicon Data index reported by Fortune. More employees, AI agents that take many steps, longer reasoning and expensive models used for simple tasks all add up.

Will AI become cheaper?

The price per unit will very likely keep falling, but total bills usually rise. Executives surveyed by Accenture in July 2026 expect token prices to fall 19% over 24 months while consumption rises 78%, which means a roughly 44% higher bill unless companies optimize.

Is it worth paying for AI if costs keep rising?

It is worth it when you can measure what each dollar buys, such as cost per document processed or per hour of staff time saved. It is not worth it as an open-ended experiment with no owner. Only 35% of large companies in Accenture's survey can calculate cost per outcome for their main AI use, which is where to start.

What is the fastest way to reduce an AI bill?

Send simple tasks to cheaper models. Accenture found 54% of AI requests go to a higher tier than needed, frontier models cost 10 to 20 times more per token than mid-tier ones, and production teams reported 40% to 60% savings from routing work to the right model.

The bottom line

AI is expensive for the same reason electricity was expensive for the first factories that electrified everything: it became cheap enough to use everywhere, so they did. That is not a reason to slow down. It is a reason to manage AI spend the way you manage any other operating cost, with a unit price you understand, a result you can measure and a person who owns both.

If you want to see what a specific workflow should cost to run, our AI automation cost guide walks through how we estimate build and running costs, and the AI ROI calculator will give you a rough payback figure in two minutes. If your AI bill has already surprised you, book a 30-minute call and we will look at where it is going. For the other side of the equation, read our breakdown of what AI automation agency pricing actually gets you.