Token Prices Fell 67%. So Why Did Your AI Bill Triple?

Key Takeaways

• Price per token is falling fast. One vendor's dataset puts the blended drop at 67% in a year. That number is real, and it is also the least useful number on your invoice.
• Agents multiply usage. A single agent run burns several times the tokens of a chat, and multi-agent setups burn more. Volume grows faster than price falls.
• Most companies find out late. Nearly half of enterprises in one recent survey say AI spend is over plan, and a quarter of firms in another miss their forecasts by more than 50%.
• Your bill has three dials: price, volume and waste. Price is the one you control least and the one everyone watches.
• Usage analytics comes first. You cannot route, cap or cut what you cannot attribute to a team, a workflow and an outcome.

Here's what doesn't add up. Every headline says tokens are cheap and getting cheaper. AI.cc's 2026 AI API Infrastructure Report says enterprise token costs fell 67% year over year, and the coverage of that report puts the blended cost per million tokens at $6.07, down from $18.40. (One caveat worth saying out loud: that number comes from one vendor's view of traffic through its own platform. I believe the direction. I would not build a budget on the decimal.)

So if the unit price fell by two thirds, why does it feel like your bill tripled?

Because your bill is price times volume, plus whatever you waste along the way. And the volume side of that equation has changed more in the last year than the price side has. Teams went from asking a chatbot questions to running agents that plan, call tools, read the results, and try again. Every one of those steps costs tokens.

Here's the mental model I've been using with the teams we work with: cheaper tokens make it cheaper to use a lot more AI, and your bill follows the "a lot more" part. Whether that is a good trade depends entirely on whether you can see where the tokens go.

The Price Per Token Is the Wrong Number to Watch

Cost per million tokens is the number vendors compete on, so it is the number that shows up in headlines. It is also the number you control least.

Think about how a finance team experiences this. Nobody on the leadership team signs off on "tokens." They sign off on a monthly invoice, a seat count, and maybe a pilot budget. When the invoice goes up, the first question is "who did this?" and in most companies nobody can answer it. The provider dashboard shows total spend by day. It does not show which team, which workflow, or which agent drove the spike.

A falling unit price also hides a behavior change. When something gets cheaper, people use more of it. That is the point. You want your team to use AI more. The problem is that "more" grows on its own schedule, and the unit price drop gets eaten in a quarter.

Track cost per task before cost per token

The number that matters is what one finished piece of work costs. One contract reviewed. One support ticket resolved. One pull request merged. If cost per task is falling while total spend rises, you are buying more output, which is fine. If cost per task is rising, something in the workflow is burning tokens without producing more value, and you need to find it.

Most teams cannot calculate this today because their usage data stops at the API key.

Agents Multiply Usage Faster Than Prices Fall

The shift from chat to agents is where the math breaks.

A chat interaction is one question and one answer. An agent is a loop. It reads its instructions, plans, calls a tool, reads the output, decides what to do next, and often rereads everything it has gathered so far before each step. Each pass through that loop is another bill.

Anthropic published a useful data point here. In its write-up on how it built its multi-agent research system, the team reports that agents typically use about 4x the tokens of a chat interaction, and multi-agent systems use about 15x. The same post says token usage alone explained most of the performance difference in their evaluation. More tokens bought better answers.

That last part is the bit people skip. The extra spend is often worth it. An agent that does a two-hour research task for a few dollars is a great deal. But it means your usage curve is now tied to how many agents your team runs, how often, and how well they are built. None of that shows up in the price per token.

Assume every loop costs something

When you review an agent workflow, ask three plain questions. How many model calls does one run make? How much context does it carry into each call? What happens when it fails and retries? A workflow that retries five times on a bad input costs six times as much as it should, and nobody notices because the output still arrives.

Budgets Break Before Anyone Notices

If this sounds theoretical, the survey data says otherwise.

In Futurum's 2H 2026 survey of 1,636 enterprise technology decision makers, 46.9% said their AI spend is running over plan. Another 10% said they have no formal AI budget to measure against at all. A separate study from Mavvrik, reported by CFO Dive, found nearly one in four companies miss their AI cost forecasts by more than 50%. (Mavvrik sells AI cost management, so read that one with the sponsor in mind. The Futurum number tells the same story from a larger, neutral sample.)

I don't read those numbers as companies being careless. I read them as companies budgeting for the chatbot era and then deploying agents. The forecast was built on last year's usage pattern. The usage pattern changed.

Put the number in front of the people making the calls

Overruns happen when the people who drive usage never see the cost. Engineers pick models. Team leads approve agent pilots. Finance sees the total a month later. Close that loop. A team lead who can see their team's spend next to their team's output will make better calls than any central policy will.

Your AI Bill Has Three Dials, and Price Is the Smallest

Here's the framework we use when a company tells us their AI spend is out of control. There are three dials, and each one needs a different fix.

Price is the dial the market moves for you, and it is already moving in your favor. Volume is mostly a good thing if it tracks to value. Waste is where the money leaks, and waste is invisible without usage data.

Find which dial moved

When the bill jumps, the first job is diagnosis. Did a new team come online? Did someone switch a workflow to a bigger model? Did an agent start looping on a bad input? Each answer points to a different fix. Cutting seats when the real problem is one runaway agent is the AI version of a hiring freeze to fix a broken process.

Visibility Comes Before Optimization

I'll say this flat: you cannot optimize AI spend you cannot attribute. Every cost-cutting idea (model routing, caching, prompt trimming, seat reassignment) depends on knowing where the tokens go today.

The tooling here is still immature, and I want to be honest about that. Provider dashboards show total usage. They don't answer business questions. Mapping a token back to a team, a project and a workflow takes setup work. Most companies skip it because the first few months of AI spend are small enough to ignore. By the time the number matters, there is no history to analyze.

The fix is boring. Instrument usage from day one. Tag it by team and workflow. Give every metric a definition and an owner. Review it monthly. That is the whole idea behind PressW's Token & Usage Analytics service: tie spend to the team, project and workflow that generated it, flag idle seats, set budgets and alerts by team, and end each monthly review with specific actions.

Instrument from day one

If you are about to roll out Claude Code or an agent pilot, set up attribution before the first user logs in. It costs almost nothing up front, and it is very hard to reconstruct later. The question you want to answer in month three is "which workflows are worth their cost?" You can only answer it if you started counting in month one.

Usage Analytics Should Show Value, Too

The worst outcome of an AI cost scare is a blanket cut. Leadership sees the invoice, panics, and caps usage across the board. The heavy users, who are often the people getting the most out of these tools, get throttled along with the idle seats.

Good usage analytics show both sides. Which teams are active, which skills and workflows run most, and what they produce. That lets you have a real conversation: this team spends a lot and ships a lot, so fund it. This team has 12 seats and three active users, so reassign them. This agent costs a couple of dollars a run and saves a person two hours, so keep it.

The tech is relatively easy nowadays. It's the people side. Getting teams to look at their own usage, own their own number, and change how they work is the actual project.

Pair every spend line with what it produced

Every row in your AI cost report should have an outcome next to it. Tasks completed, tickets closed, hours saved, whatever fits the workflow. Spend without outcome is just a number someone will eventually cut.

Cheap Tokens Are a Gift If You Can See Them

I'm bullish on where this goes. Falling prices mean more workflows clear the bar for automation every quarter. That is good for everyone building with AI.

But I don't have a clean answer for how fast usage will grow from here, or what the right budget model looks like for agents that run on their own. Per seat? Per workflow? Per outcome? I think most companies, including the ones we work with, are still figuring that out. What I am sure of is that the companies that can see their usage will figure it out first.

If your AI bill surprised you this quarter, start with one question: can you name the three workflows that drove most of it? If you can't, that is the first thing to fix. We help teams set that up through Token & Usage Analytics, and if you want to compare notes on what you're seeing, Book a Discovery Call.

FAQ