Why AI Token Costs Are Changing in 2026 (and What It Means for Your Business)

Here's a contradiction a lot of business owners are running into this year: every headline says AI is getting cheaper, and yet the AI bill on your desk keeps creeping up. Both of those things are true at once, and once you see why, the rest of your AI budget starts making a lot more sense.

Think of it the way you'd think about gas prices. If the price per gallon drops but you start driving five times as far every week, your total fuel bill goes up even though gas is cheaper than it's ever been. AI token costs are doing exactly that right now — the price per unit is falling fast, but businesses are using so much more of it that the total spend climbs anyway.

The short version: the price per token is falling. Your total bill depends on price and usage — and usage is growing faster than the price is dropping.

This piece walks through why token prices are dropping, why that hasn't translated into a lower bill for most companies, and what any business — not just ones with a dedicated finance team watching every line item — can actually do about it.

The confusing headline: token prices are collapsing

Start with the part of the story most coverage already gets right: AI token prices really have fallen off a cliff. GPT-4o's input pricing, for instance, dropped from $5.00 to $2.50 per million tokens within about five months of its May 2024 launch. Look at blended costs across providers and the picture is similar — the average cost per million tokens fell from roughly $18.40 to $6.07 between the first quarter of 2025 and the first quarter of 2026. Depending on which model tier you're comparing, that's somewhere in the neighborhood of a 70 to 90 percent price decline since 2023 and 2024.

If you only read the pricing pages, you'd assume your AI costs should be shrinking right along with them. For a lot of businesses, they aren't.

What a token actually is (and how you're billed for it)

Before going further, here's exactly what you're paying for. A token is roughly four characters or about three-quarters of a word in English — a 500-word email works out to somewhere around 650 tokens. Every request you send an AI model gets broken into these small chunks, and providers bill you for two separate streams: the tokens you send in (your prompt, any documents or context attached to it) and the tokens the model sends back (its response).

Output tokens almost always cost more — typically three to five times the input rate. The reason comes down to how the model actually works: it can read your input more or less all at once, but it has to generate its response one token at a time, in sequence. That's more computational work per token, so it costs more per token. It's a small detail, but it matters once you start looking at where your bill actually comes from — a workflow that asks for a short answer costs very differently than one that asks the model to write three pages.

Why prices are falling

The price collapse isn't a temporary promotion. It's the result of a few things happening at once. Model providers have gotten much better at building smaller, more efficient models that deliver similar quality for a fraction of the compute — techniques like distillation and better inference engines mean a model doesn't need as much raw horsepower to produce a good answer as it did two years ago. At the same time, the market has gotten genuinely competitive: OpenAI, Anthropic, Google, DeepSeek, and a growing list of others are racing each other on price as much as on capability, and that kind of competition pushes prices down fast.

There's also been a deliberate shift toward cheap, lightweight model tiers — the "mini" and "flash"-class models built specifically to handle simple, high-volume tasks at a fraction of what a flagship model costs. These aren't scaled-down afterthoughts anymore; they're a real product category, and they're a big part of why the average price per token has fallen so much even as flagship models stay expensive.

Why your bill is rising anyway (the paradox)

So if prices are genuinely falling, why does the invoice keep climbing? Go back to the gas-price analogy: the price per gallon is down, but total spend depends on price and how much you're using. And AI usage isn't just growing — it's growing faster than the price is falling.

The numbers back this up in a pretty striking way. Bain & Company found that token costs were cut roughly in half between December 2024 and December 2025, while the actual volume of tokens consumed grew by about 450 percent over that same stretch. Put those two numbers next to each other and the outcome isn't subtle: even with prices cut in half, total spend goes up when usage grows nine times faster than the discount. Separately, the Silicon Data LLM Token Expenditure Index — which tracks what the market actually pays per million tokens — shows the price of a single token has dropped more than 90 percent since 2023, while overall spending on large language models has roughly doubled since late 2025.

There's actually a name for this pattern, and it's not new. In 1865, an English economist named William Stanley Jevons noticed something odd about coal: when steam engines got more efficient and needed less coal to do the same amount of work, coal consumption didn't fall — it went up. Cheaper, more efficient energy didn't reduce demand; it unlocked new uses for it that hadn't been worth the cost before. Economists now call this the Jevons paradox, and it's the same mechanic playing out with AI tokens. As the cost of "thinking" with AI drops, companies don't pocket the savings — they run more automated workflows, deploy more agents, and ask AI to do things that weren't economical a year ago. The unit price falls. The total bill doesn't.

The real drivers: agents, reasoning tokens, and retries

That's the mechanism in the abstract. In practice, two overlapping dynamics are driving most of the extra consumption.

The biggest one is agentic AI — systems that don't just answer a single question but plan, take actions, check their own work, and loop back to try again. A Stanford Digital Economy Lab study of agentic coding tasks found they can consume up to a thousand times more tokens than an equivalent code-chat exchange, because instead of one exchange, you're paying for a chain of API calls: read the file, form a plan, execute a step, check the result, retry if something's off, revise, and repeat until the task is done. Each of those calls resends much of the context that came before it. That's the extreme end of the range for coding specifically, but the same mechanic scales down to any agentic workflow — more steps and more retries mean more resent context, which means more tokens.

Reasoning models add another layer. Models built for complex, multi-step reasoning generate internal "thinking" tokens before they produce a visible answer — tokens you're billed for but never actually see. Per-token rates for these models already run 10 to 80 times higher than standard models, and on a genuinely hard problem, one documented example shows a model burning 20,000 reasoning tokens before writing a 400-token answer, meaning the real bill reflects roughly 50 times more tokens than the visible response would suggest.

This isn't hypothetical for the companies living it:

• Uber gave roughly 5,000 engineers access to Anthropic's Claude Code in December 2025. Adoption climbed from about a third of engineers to 84 percent within a few months, and by around April 2026 the company had burned through its entire annual AI budget. It has since capped spending at $1,500 per employee, per AI coding tool, per month.

• Microsoft cancelled most internal Claude Code access in one of its divisions over cost concerns about six months after rolling it out in December 2025, redirecting engineers to its own GitHub Copilot CLI instead.

• Accenture reportedly told staff to cut back on AI use for tasks that didn't need it, after its agentic AI strategy lead described "rapid escalation in AI token spend" — driven not by engineers, notably, but by non-engineering staff using AI for simple tasks like converting PDFs to slides.

These aren't small companies mismanaging a budget. They're sophisticated operations that got caught off guard by how fast agentic usage compounds.

What this means for your business

The practical takeaway is that AI spend needs to be treated as a variable operating cost, not a flat subscription line you can forecast and forget. That's a real mindset shift for a lot of teams that budgeted for AI the way they'd budget for a software license.

It also helps to know where you actually sit relative to other businesses, rather than benchmarking against enterprise-scale horror stories. Ramp, which processes AI vendor payments for thousands of businesses, found the median company spends around $2,246 a month on AI tokens — but the average is $140,842 a month. That enormous gap isn't a typo; it reflects the fact that most companies are still in early or moderate AI adoption, while a small number of heavy users — usually a team running an unchecked automated workflow — pull the average way up. If your spend looks nothing like $140,000 a month, you're probably not an outlier; you're the median.

A more useful way to think about your own budget is per employee, per month, rather than as one lump total — it normalizes for company size and gives you something to actually compare month to month. And the single most useful early-warning sign isn't the dollar amount itself, it's the trend: if your AI spend is climbing 40 percent or more month over month while headcount stays flat, that's the pattern worth digging into. Not necessarily a crisis — a signal that usage is scaling faster than anyone planned for.

Practical steps you can take now

None of this means you should pull back from AI — it means the budgeting has to catch up to how the technology is actually being used. A few things actually reduce AI token costs without cutting off access:

1. Match the model to the task. Not every request needs your most expensive, most capable model. Route simple, high-volume tasks to a lightweight model and save the frontier-tier model for the work that actually needs it.

2. Use prompt caching where it's available. If your workflows repeat the same instructions or context across many calls, caching lets you avoid paying full price to process that same text over and over.

3. Set spending alerts before they're a surprise. A soft budget per team or per employee, with an alert when it's approaching the limit, catches runaway usage while it's still a conversation instead of a crisis.

4. Get visibility by workflow, not just by invoice. A single monthly total from your AI provider tells you almost nothing about which team or feature is driving the spend. Breaking it down by workflow is what actually lets you act on it.

5. Ask whether a new agentic workflow is worth what it costs before you scale it. Agentic automation can be genuinely valuable, but it's also the single biggest lever on your bill. Pilot it, measure what it actually costs against what it actually saves, and only then decide whether to roll it out further.

FAQ

Why is my AI bill going up if token prices are falling? Because your total cost depends on price and usage together, and usage is growing much faster than prices are falling. The unit price is genuinely lower; you're just using a lot more of it — often without realizing how much a new automated or agentic workflow adds to the total.

What is a token in AI pricing? A token is the basic unit AI providers use to measure and bill for text — roughly four characters or three-quarters of a word. Providers charge separately for the tokens you send in and the tokens the model generates back, with output tokens typically costing several times more than input tokens.

What is the Jevons paradox and how does it apply to AI? It's an economic pattern, first observed with 19th-century coal use, where making a resource more efficient (and cheaper) leads to more total consumption of it, not less, because the lower cost unlocks new uses that weren't worth it before. AI tokens are following the same pattern: cheaper tokens haven't reduced spending, they've made it economical to run far more AI workflows than before.

How much should a business budget for AI per employee per month? It varies enormously by how deeply AI is embedded in your workflows, but Ramp's data puts the overall median at around $46 per employee per month across the businesses it tracks — a useful starting benchmark, though companies using AI more deeply run well above that. Track your own per-employee spend over a few months and watch the trend rather than anchoring to a single target number.

Will AI token prices keep falling in 2026 and beyond? Most signs point to unit prices continuing to fall as competition among providers stays intense and models keep getting more efficient. That said, a falling unit price is no guarantee your total bill will fall with it — that depends entirely on how much your usage grows in the meantime.

Is AI actually cheaper than hiring employees? It depends heavily on the task. For some workloads, the compute cost has genuinely started to exceed what a human employee would cost for the same output — Nvidia VP of applied deep learning Bryan Catanzaro put it plainly in an interview, saying the cost of compute for his team was "far beyond the costs of the employees." For other tasks, AI is still dramatically cheaper. The honest answer is that it's task-specific, and worth actually measuring rather than assuming either way.