The Cheaper AI Gets, the Bigger Your Bill

Key Takeaways

• The deflation is real: Querying an Artificial Intelligence (AI) model at GPT-3.5-level performance fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024, a more than 280-fold collapse.
• The bills went up anyway: 98% of organizations surveyed by the FinOps Foundation now actively manage AI spend, up from 31% just two years ago.
• June 2026 was the month the flat fee died: GitHub moved every Copilot plan to metered, token-based billing on June 1. Anthropic announced a similar split, then paused it on June 15, the day it was due to take effect.
• The culprit is volume, not price: Agentic workloads consume orders of magnitude more tokens per task than a chat query. Cheaper units multiplied by exploding consumption equals a bigger invoice. That is the Jevons paradox, running at data-center scale.

The Month the $20 Plan Stopped Making Sense

For three years, the deal was simple: hand an AI vendor a flat monthly fee and use the model like an all-you-can-eat buffet. In June 2026, the buffet closed.

On June 1, GitHub replaced Copilot’s “premium requests” with GitHub AI Credits across every plan, with usage consumed “based on token usage, including input, output, and cached tokens, according to the published API rates for each model.” Two weeks later, Anthropic was scheduled to push through its own version: moving Claude Agent SDK (Software Development Kit) and third-party agent usage off subscription pools and onto separate monthly credits, ranging from $20 on the Pro plan up to $200 on the top tier. It never happened. On June 15, the day the change was due to take effect, Anthropic paused it, saying “Nothing changes for now” and that it was “working to better align the plan with actual usage patterns.”

Here is what makes this wave of repricing genuinely strange: it arrived at the exact moment AI tokens became the cheapest they have ever been. Understanding why both things are true at once explains where your AI bill is actually headed.

The Great Deflation Is Real

Start with the vendors’ side of the story. It has the advantage of being true.

Stanford’s AI Index, the most-cited annual accounting of AI progress, found that the cost of querying a model that scores the equivalent of GPT-3.5 on MMLU (Massive Multitask Language Understanding, a standard knowledge benchmark) “dropped from $20.00 per million tokens in November 2022 to just $0.07 per million tokens by October 2024,” a more than 280-fold reduction in approximately 18 months. Depending on the task, the same report found that inference prices have fallen anywhere from 9 to 900 times per year. Underneath the software, machine learning hardware costs have dropped about 30% per year while energy efficiency improved roughly 40% annually.

A price collapse of that magnitude should make AI budgets a rounding error. Instead, the opposite happened.

The Consumption Explosion

The FinOps Foundation, the industry body for cloud Financial Operations (FinOps), found in its State of FinOps 2026 report that 98% of organizations now manage AI spend, based on the 693 practitioners who answered that question, up from 63% in 2025 and 31% in 2024. AI cost management ranked as the number one skillset teams need to develop. You do not build an entire discipline around a cost that is shrinking.

What changed is not the price of a token. It is how many tokens a single piece of work consumes.

A 2023-era chatbot exchange was one request: a question in, an answer out, a few thousand tokens end to end. An agentic coding session is a different animal entirely. The agent reads your repository, plans, calls tools, reads their output, fails, retries, and verifies its own work. Each of those steps is a fresh model call carrying the accumulated context of everything before it. A four-second autocomplete and a forty-five-minute autonomous refactor sit at opposite ends of a token gap that spans roughly three orders of magnitude.

Gartner put a date on where this leads. In late June 2026, the research firm forecast that spending on AI coding agents is on track to surpass the average software developer’s salary by 2028, reporting that nearly one-quarter of technology leaders already spend between $200 and $500 per developer each month on AI coding tokens, and about 6% report spending more than $2,000 per developer monthly. Whether or not that specific forecast lands, the direction is not in dispute. The unit got cheap; the workload got enormous.

Anatomy of a Quiet Repricing

The flat-fee subscription was priced for the chatbot era, and the agent era broke its math. What followed was less a coordinated price hike than a series of increasingly awkward corrections.

GitHub went first and went furthest. Under the new Copilot scheme, plan prices stay the same, but each plan now includes a matching allowance of AI Credits: Pro at $10 per month includes $10 in credits, Business at $19 per user includes $19, Enterprise at $39 includes $39. One credit is worth $0.01. Code completions and Next Edit suggestions remain unmetered, so the casual autocomplete user notices nothing. The agent user, in contrast, is now spending down a token meter running at published API (Application Programming Interface) rates.

Anthropic tried, then blinked. The plan would have kept chat and the official Claude Code command line interface (CLI) on standard subscription limits, while stopping the Agent SDK, the headless claude -p command, and third-party apps from drawing on those limits, billing them at prevailing API rates instead, cushioned by those monthly credits. The pause on June 15 came as the Wall Street Journal reported that OpenAI is considering steep price cuts for its own API, which would have made unilaterally raising effective prices an uncomfortable move. A repricing that one vendor ships and its rival abandons on launch day tells you the industry has not yet found the floor.

The subtler lever is the tokenizer. Anthropic’s own pricing documentation notes that Claude Opus 4.7 and later models use a tokenizer that “produces approximately 30% more tokens for the same text.” Same list price, same text, more billable units. The site flagged exactly this dynamic when Opus 4.7 launched at unchanged headline pricing.

And list prices themselves have quietly stopped falling at the frontier. Anthropic’s price sheet shows Claude Sonnet 5 at an introductory $2 per million input tokens and $10 per million output tokens through August 31, 2026, after which standard pricing of $3 and $15 takes effect, a scheduled 50% increase. The new top-tier Claude Fable 5 lists at $10 in and $50 out, double the $5 and $25 of Claude Opus 4.8. Even the budget rung moved up: Claude Haiku 4.5 lists at $1 and $5, above the $0.80 and $4 of the retired Haiku 3.5 it replaced. On this one price sheet, every tier is repricing upward. The deflation that remains lives in the competition, in cheaper rivals and open-weight models, not in the incumbent’s list prices.

The Data

The whole paradox fits in one line of arithmetic. What a task costs is the price of a token times the number of tokens the task eats:

$$C_{task} = P_{token} \times T_{task}$$

Between 2022 and 2024, $P_{token}$ fell roughly 280-fold for constant quality. But for the workloads that now dominate, $T_{task}$ grew by a factor of hundreds to thousands. When the second factor grows faster than the first one shrinks, $C_{task}$ rises even as every individual token gets cheaper.

Run the numbers on the new Copilot plan to see how fast a meter drains. Take an agentic session on a frontier model priced like Claude Opus 4.8, at $5 per million input tokens and $25 per million output tokens. A session that processes 1.5 million input tokens and generates 100,000 output tokens costs exactly $10.00 at those rates:

Two sessions like that and a Copilot Business seat has burned through its entire $19 monthly credit allowance. Everything after that is metered overage. The flat fee did not get more expensive. It just stopped covering the way people actually work now.

The False Floor

So who is the villain here? Pick your theory.

One read is cynical: the flat fees were subsidized land grabs, and June was the clawback, the moment the subsidized floor started to break. The other read is boring: nobody needed to plan anything. A cheap flat fee set against a workload that can consume hundreds of dollars of compute at API rates is arithmetic that fails on its own, no malice required.

The evidence favors the boring read, with a twist. This is not a supply squeeze. The same week Anthropic’s pause settled in, Reuters reported that Meta plans to sell its excess AI computing capacity through a cloud business, per Bloomberg News. Capacity is not scarce; frontier-scale consumption is just outrunning the pricing models built for chat.

And customers have somewhere to go. Reuters reported on July 2 that a new, inexpensive Chinese AI model is catching up with Anthropic’s and OpenAI’s frontier offerings. Tom’s Hardware has documented firms shifting workloads to Chinese and open-weight models specifically to stretch AI budgets as subscriptions hit a pricing wall. Every metered invoice is an advertisement for the cheaper model one API call away. That, more than customer goodwill, is the check on how far the repricing can go.

The 1865 Rhyme

None of this is new. In 1865, the economist William Stanley Jevons observed that as coal-burning technology became more efficient, Britain’s total coal consumption rose rather than fell, because efficiency made coal-powered industry economical in far more places. Cheaper units, more total burn. Swap coal for tokens and the AI industry is running the same experiment with the same result.

The nearer rhyme is fiber. In the late 1990s, telecom carriers borrowed enormous sums to build out capacity that grew far faster than demand, and when financing dried up after the dot-com bubble burst, their debt loads drove a wave of bankruptcies that took down giants like WorldCom and Global Crossing. The lesson from that era is not that demand was fake. Demand kept growing; it simply grew slower than the infrastructure and the business models built to serve it, and the correction arrived all at once. The economics of AI video generation already hit this wall; text agents are hitting it in slow motion.

What Happens Next

The near-term trajectory is visible in the June wreckage.

Metering spreads. GitHub has already shown that the politically survivable version is “plan price unchanged, agent usage metered,” and the pause at Anthropic reads as a revision, not a retreat; the company said it is realigning the plan, not abandoning it. Watch September 1, when Sonnet 5’s scheduled 50% list-price increase takes effect and tests whether mid-tier prices can actually rise without losing workloads to open-weight alternatives.

Expect the routing layer to become the interesting battleground. Firms are already pairing frontier models with alternatives that cost roughly ten times less for routine work. Once every token is metered, sending the easy majority of requests down the cheap path stops being an optimization and starts being the whole game. And expect “token FinOps” to become a job title. When 98% of organizations are managing AI spend and the number one skill gap is AI cost management, someone is getting hired to own the meter.

What This Means for You

If you write code with AI tools:

• Check which of your tools moved to metered billing in June. The autocomplete tier is still effectively flat; the agent tier is not.
• Match the model to the task. Running a frontier model on work a cheap model handles is now a visible line item, and the price gap between tiers is an order of magnitude.

If you own an AI budget:

• Forecast on tokens per task, not seats. Seat-based budgeting is how 2026’s overruns happened.
• Treat vendor list prices as the start of the math, not the answer. Tokenizer changes and consumption growth move real costs even when the price sheet doesn’t.

Frequently Asked Questions

If token prices keep falling, won’t this fix itself?

Unit prices probably will keep falling; Stanford’s data shows declines of 9 to 900 times per year depending on the task. But consumption per task has grown faster than prices have fallen, and agents are still getting more autonomous. Betting your budget on deflation outrunning your own usage has been a losing bet for two straight years.

Did Anthropic cancel its billing change or just delay it?

Paused, not canceled. The company said “Nothing changes for now” and that it is working to realign the plan with actual usage patterns. The structural pressure that produced the change did not go away.

Is this just AI vendors price-gouging?

The evidence points the other way. Metered billing at published API rates is more transparent than a flat fee that quietly subsidized heavy users, and competition from inexpensive Chinese and open-weight models limits how far anyone can push. The real story is that the unit of pricing (a month of access) no longer matched the unit of cost (a token), and the mismatch finally broke.

The Bottom Line

The era of cheap tokens and the era of giant AI bills are the same story told from opposite ends of the invoice. Deflation made intelligence cheap enough to spend recklessly, agents industrialized the spending, and the flat-fee subscription that hid the meter died of arithmetic in June 2026. The correction now underway is uncomfortable but honest: visible meters, real prices, and a market where the cost of a task finally tells you the truth about what it took to run. Budgets built on that truth will do fine. Budgets built on the buffet staying open will not.