Dev.to AI ๐Ÿค– Ai ๐Ÿ‘ 0 ๐Ÿ“– 16 min read

AI Tokenomics: Why Cheaper Tokens Are Making Your AI Bill Bigger

The price of a given level of AI capability is falling by roughly an order of magnitude every year. Yet companies are blowing through AI budgets, flat-rate AI plans are disappearing, and token spending keeps climbing. Th

The price of a given level of AI capability is falling by roughly an order of magnitude every year. Yet companies are blowing through AI budgets, flat-rate AI plans are disappearing, and token spending keeps climbing. This is not a pricing anomaly. It is the economics of tokens working exactly as economics predicts. Here is what is happening on the supply side, the demand side, and in your budget, and the controls that actually keep AI spend under control.

Research checked November 21, 2026.

The Budget Was Gone by April

In June 2026, Fortune reported a detail that should make every engineering leader pause. Uber's chief operating officer said the company had used up its entire AI budget in the first four months of the year. Bloomberg later reported that Uber capped AI spending at $1,500 per employee per month.

Uber is not a company that misunderstands technology costs. And the tokens it was buying were getting cheaper the whole time.

The same Fortune piece cited the Silicon Data Token Expenditure Index: the price of a single token has dropped more than 90 percent since 2023, while spending on large language models has doubled since late 2025. Bain and Company analysts put it more bluntly in a brief quoted in the article:

"The models get cheaper. The usage gets heavier. The bill stays stubbornly high."

This looks like a paradox. It has a name, and it is more than 150 years old.

In 1865, the economist William Stanley Jevons observed that more efficient steam engines did not reduce Britain's coal consumption. They increased it. Cheaper energy per unit of work made coal worth using for far more work.

Apollo's chief economist, Torsten Slok, applied the same idea to AI this year: "As tokens get cheaper, companies don't spend less but instead run more AI agents, automate more workflows."

That is AI tokenomics in one sentence.

Falling token prices do not shrink AI budgets. They expand what is worth automating. Vendors set the price of a token. Your architecture, your agents, and your incentives set how many tokens you use. The organizations that control AI spend will be the ones that govern demand, not the ones waiting for prices to fall.

TL;DR

  • Token prices are collapsing at fixed capability. Epoch AI estimates the cost of a given level of AI performance has fallen about 47 percent per quarter since 2023, or about 13x per year. MIT researchers estimate 5x to 10x per year for frontier-level benchmark performance.
  • Frontier costs are not falling the same way. The same MIT paper estimates the price of running frontier models is rising 3x to 18x per year, because of bigger models and larger reasoning demands.
  • Token volume is growing faster than prices fall. Google said at I/O 2026 that it processes 3.2 quadrillion tokens per month, up from 9.7 trillion two years earlier. Its model APIs alone reached nearly 32 billion tokens per minute in Q3 2026, more than four times a year earlier.
  • Spending is the product of both. Bain found token costs halved from December 2024 to December 2025 while tokens consumed grew 450 percent. Half the price times 5.5x the volume is 2.75x the spend.
  • Agents and reasoning are the multiplier. Anthropic reported that agents use about 4x more tokens than chat, and multi-agent systems about 15x. Reasoning tokens, longer contexts, and background agents push the number higher.
  • Cheaper tokens unlock new workloads. Each price drop moves more tasks above the break-even line, so total demand grows. This is the Jevons effect, and AI demand appears to be highly elastic.
  • Flat-rate AI is breaking. Pricing designed for chat cannot absorb agents that run for hours. Anthropic introduced weekly limits for Claude Code in 2025, and other AI coding tools have moved toward metered or capped pricing.
  • For AI product builders, demand is margin risk. If you charge per seat and your users adopt agents, your costs can grow faster than token prices fall.
  • For buyers, forecast volume, not price. A budget that assumes token prices will fall is the budget that runs out in April.
  • Govern demand. Budgets per agent and per employee, routing by task difficulty, caching, reasoning-effort tuning, loop caps, and cost attribution are the levers that work.

First, Separate the Prices of a Token

"Tokens are getting cheaper" and "AI is getting more expensive" are both true. They describe different numbers.

Number Direction What drives it Who controls it
Price per token at fixed capability Falling fast Better hardware, better algorithms, competition Vendors
Price per token at the frontier Flat or rising Bigger models, more reasoning, premium tiers Vendors
Tokens per task Rising Agents, reasoning, longer context, retries Your architecture
Tasks per month Rising New use cases, adoption, background automation Your organization
Total spend Usually rising All of the above multiplied together Mostly you

The core equation is simple:

AI spend  =  price per token  x  tokens per task  x  tasks

Vendors push the first term down. Almost everything about modern AI adoption pushes the second and third terms up.

When people say AI is getting cheaper, they are usually looking at the first row. When finance says AI is getting more expensive, they are looking at the last.

The Supply Side: Why Tokens Keep Getting Cheaper

The decline in token prices is one of the fastest cost curves in modern technology.

The measurements

Epoch AI published "The plunging price of thought" in September 2026. Its estimate: "The cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13ร— per year."

The decline is not uniform over time. Averaging across five benchmarks, Epoch found cost falls "66% per quarter (75ร— per year) for performance that has just debuted as SOTA," and "two years later, prices fall half as fast, at 32% per quarter (4.7ร— per year)." New capability gets cheap fastest right after it arrives, then the decline slows.

a16z's Guido Appenzeller coined the term "LLMflation" in 2024: "For an LLM of equivalent performance, the cost is decreasing by 10x every year." He also cautioned that "its rate may slow down."

MIT researchers in "The Price of Progress" estimate that the price for a given level of benchmark performance has decreased "around 5ร— to 10ร— per year" for frontier models across knowledge, reasoning, math, and software engineering benchmarks. After removing the effects of competition and hardware price declines, they estimate algorithmic efficiency progress alone at "around 3ร— per year."

Different methods, different benchmarks, same conclusion: at a fixed level of capability, tokens get dramatically cheaper every year.

Tokens as a manufactured product

The supply side is also being reframed in industrial terms. NVIDIA now describes data centers as "AI factories," and its own materials treat tokens per megawatt as the measure that governs a factory's earning capacity. Power, not chips alone, is increasingly the constraint, so efficiency per watt becomes efficiency per dollar.

That framing matters for buyers. Tokens are becoming a manufactured commodity with a production cost, and production costs keep falling with each hardware generation.

The part that does not get cheaper

There is an important caveat in the MIT paper. While the price of a given level of performance falls, "the price of running frontier models is rising" between 3x and 18x per year "due to bigger models and larger reasoning demands."

That is the first half of the tokenomics paradox. Last year's capability gets cheap. This year's capability costs more. And most teams upgrade.

The Demand Side: Why Token Volume Is Exploding

If prices tell half the story, volume tells the rest.

Google's numbers

Google publishes some of the clearest public token data.

At Google I/O 2026 in May, Sundar Pichai described the growth in tokens processed across Google's surfaces: 9.7 trillion tokens per month two years earlier, about 480 trillion at I/O 2025, and "3.2 quadrillion tokens per month" at I/O 2026.

That is roughly a 330x increase in two years.

The API-only numbers, reported in Alphabet earnings remarks, show the same curve:

Period Google first-party model APIs
Q1 2026 About 16 billion tokens per minute
Q2 2026 (July 22) About 22 billion tokens per minute
Q3 2026 (October 28) Nearly 32 billion tokens per minute, "more than four times what we processed a year ago"

Google also reported internal developer usage rising from half a trillion tokens a day in March 2026 to more than 3 trillion a day by I/O in May.

These figures use different scopes and units, so they should not be stacked into one series. Together they show something unambiguous: token demand is compounding much faster than token prices are falling.

The enterprise numbers

Bain's analysis, as quoted by Fortune, found that token costs halved from December 2024 to December 2025, while tokens consumed grew 450 percent over the same period. Bain attributed part of the growth to companies upgrading to newer models rather than keeping existing ones and banking the savings.

Run the arithmetic:

price:   x 0.5
volume:  x 5.5   (450% growth)
spend:   0.5 x 5.5  =  x 2.75

Prices halved. The bill almost tripled.

What is driving demand

Five forces are multiplying tokens per task and tasks per month at the same time.

1. Agents. Anthropic reported in 2025 that "agents typically use about 4ร— more tokens than chat interactions, and multi-agent systems use about 15ร— more tokens than chats." Every workflow that moves from chat to agent multiplies its token footprint.

2. Reasoning. Reasoning models spend tokens thinking before they answer. Those tokens are billed, even when you do not see them.

3. Longer context. Agents re-read growing histories, tool results, and documents on every step.

4. Background and proactive work. Agents that monitor inboxes, run scheduled routines, or work overnight consume tokens without a human prompt. As I wrote about Claude Duet, an agent that works while you sleep also spends while you sleep.

5. Upgrades. When a better model ships, teams switch to it, as Bain noted. The price drop on last year's model never reaches the budget.

The Jevons Mechanism in AI

The Jevons paradox happens when demand is elastic enough that a lower price increases total spending. In AI, three feedback loops make demand very elastic.

Loop 1: New use cases cross the break-even line

Every task has a value and an AI cost. When AI cost falls below the task's value, the task becomes worth automating.

worth automating  if  value of task  >  tokens per task x price per token

A tenfold price drop does not make existing tasks ten times cheaper and leave it there. It moves a whole new set of tasks above the line: summarizing every support call instead of a sample, reviewing every pull request instead of large ones, monitoring every log stream instead of alerts only. Each new category adds volume.

Loop 2: Savings are reinvested in quality

When a cheaper model can do last year's work, teams rarely pocket the savings. They use the headroom to run a stronger model, add a verification step, increase reasoning effort, or run a second agent to check the first. Quality improves. The token bill does not shrink.

Loop 3: Autonomy scales consumption

Cheaper tokens make longer agent runs affordable. Longer runs make agents more capable. More capable agents get assigned bigger tasks. Fortune's summary of the trend captures it: once a team believes AI can handle bigger tasks, it asks for more and uses more tokens.

Slok summarized the result: "Lower cost per interaction does not mean fewer interactions."

Who Pays: The End of Flat-Rate AI

Flat-rate AI pricing was designed for chat. A person types, the model replies, the person reads. Usage is bounded by human attention.

Agents removed that bound.

An agent can run for hours, spawn sub-agents, and work overnight. A single heavy user running agents can consume many times the tokens of a typical chat user on the same plan.

The industry is adjusting:

  • Anthropic introduced weekly rate limits for Claude Code subscribers in 2025, citing users who ran it continuously and saying the change would affect a small minority of users.
  • Other AI coding tools, according to 2026 trade coverage, have moved from unlimited plans toward token-based credits, metered billing, or tighter caps.
  • Enterprises are introducing their own caps, like Uber's reported $1,500 per employee per month.

None of this means AI is becoming unaffordable. It means the cost of agentic usage is becoming visible, and someone has to own it: the vendor, the employer, or the user.

What Tokenomics Means If You Build AI Products

If you sell an AI product, token demand is not just your customer's problem. It is your gross margin.

Here is an illustrative example. The numbers are assumptions, but the structure is common.

Year 1. You charge $30 per user per month. Most users chat; a few run agents.

User type Share of users Tokens per month Blended price per 1M tokens Token cost
Chat user 90% 2M $1.50 $3
Agent user 10% 60M $1.50 $90
Average $11.70

Average token cost is $11.70 per user. Gross margin on tokens is about 61 percent.

Year 2. Token prices halve. Good news. But 40 percent of users now run agents, and agent workflows have doubled in size.

User type Share of users Tokens per month Blended price per 1M tokens Token cost
Chat user 60% 2M $0.75 $1.50
Agent user 40% 120M $0.75 $90
Average $36.90

Average token cost is now $36.90 against $30 of revenue. Token prices halved, and gross margin went from about +61 percent to about -23 percent.

That is the tokenomics trap for product builders. Your costs are driven by your users' behavior, and agents change user behavior faster than vendors cut prices.

What I would design in from day one:

  • Metering everywhere. Know tokens per user, per feature, and per workflow.
  • Pricing that tracks usage or value. Seats plus usage, credits, or outcome-based pricing, rather than unlimited seats.
  • Model routing. Send easy work to cheap models, and keep frontier models for tasks that need them.
  • Caching and compaction. Reuse stable context instead of paying for it on every call.
  • Fair-use caps with clear communication. Limits users discover by surprise damage trust more than limits they understand.

What Tokenomics Means If You Buy AI

For engineering and finance leaders, the lesson is about forecasting.

Forecast volume, not price. A budget built on the assumption that token prices will fall 50 percent is likely to be wrong in the dangerous direction. Prices may fall. Usage will almost certainly rise faster.

Attribute every token. Spend that cannot be assigned to a team, product, workflow, or agent cannot be managed. Tag model calls with the owner and purpose.

Measure value per token, not tokens. In my previous article, I argued for cost per verified successful task as the core unit metric. Tokenomics is the organizational version of the same idea: total spend only matters relative to the value it produces.

Treat token budgets as product decisions. Deciding that an internal coding agent gets $500 per engineer per month, or that a support agent may spend up to $2 per resolved ticket, is a decision about what work is worth automating. It belongs with product and engineering leaders, not only finance.

Demand Governance: The Controls That Actually Work

If vendors control price and you control demand, these are the levers worth building.

1. Budgets at the right level

Set budgets per agent, per task, per team, and per employee, and enforce them in code, not in a spreadsheet. A task-scoped budget, like the spend envelopes I described for agent payments, stops runaway runs before they become monthly surprises.

2. Route by task difficulty

Most tasks do not need the frontier model. Route classification, extraction, and routing decisions to small models or decision models such as Jev, and reserve frontier models for work that measurably benefits.

3. Tune reasoning effort per step

Princeton's Holistic Agent Leaderboard found higher reasoning effort reduced accuracy in the majority of its runs. Reasoning is a parameter to measure, not a dial to turn up by default.

4. Cache and compact

Cache stable prefixes such as system prompts, tool definitions, and reference documents. Compact long histories. These cut tokens without cutting quality.

5. Cap loops and tails

Set maximum steps, tokens, and wall-clock time per task. A small number of runaway agents can dominate a monthly bill.

6. Make background work earn its keep

Scheduled and proactive agents run without a human asking. Review what they produce. A daily briefing nobody reads is pure token spend.

7. Show people their usage

Engineers and teams make better decisions when they can see what their agents cost. Visibility changes behavior faster than policy documents do.

What Could Change the Curve

Tokenomics is a trend, not a law. Several things could bend it.

Price declines may slow. Epoch's own data shows the decline slowing as capability matures, and both a16z and Epoch warn against assuming the current pace continues indefinitely.

Power may set a floor. If electricity and data center capacity become the binding constraint, tokens per megawatt matters more than chip price, and token supply may not keep pace with demand.

Frontier costs may keep rising. If the best models keep getting bigger and reason longer, staying at the frontier will cost more each year, even as older capability becomes nearly free.

Subsidies may unwind. Some AI pricing has been set to win market share rather than to cover cost. The shift away from flat-rate plans suggests that correction has started.

Demand may saturate. At some point, organizations will have automated the obvious work. That point does not appear close.

I would not bet on any of these reversing the core dynamic in the next two years. I would plan for token demand to keep growing faster than token prices fall.

A Practical 30-Day Plan

Week 1: Baseline

  • Pull token usage and spend by model, team, product, and workflow.
  • Identify the top 10 workloads by spend.
  • Separate chat usage from agent and background usage.
  • Find any spend you cannot attribute to an owner.

Week 2: Unit economics

  • For the top workloads, calculate tokens per task and cost per successful outcome.
  • Identify which workloads use frontier models for work a smaller model could handle.
  • Measure what share of spend comes from the top 1 percent of tasks.

Week 3: Controls

  • Add budgets and caps for the highest-spend agents.
  • Route one high-volume workload to a cheaper model and compare outcomes.
  • Enable caching for stable prompt prefixes.
  • Review background and scheduled agents for value.

Week 4: Forecast

  • Build next year's forecast from expected task volume and tokens per task, not just expected price changes.
  • Model a scenario where agent adoption doubles.
  • Set budget owners and a monthly review.

At the end of 30 days, you should know where your tokens go, which ones produce value, and how fast demand is growing. That is more useful than any price prediction.

Seven Questions for Any AI Budget Owner

  1. What share of our token spend comes from agents and background work?
  2. How fast is token volume growing per team, and why?
  3. Which workloads use frontier models, and do they need to?
  4. What is our cost per successful outcome for the top five workloads?
  5. Which spend cannot be attributed to an owner?
  6. What happens to our budget if agent adoption doubles?
  7. If we sell AI features, what does our heaviest 10 percent of users cost us?

If your forecast assumes prices will fall and usage will stay flat, it is not a forecast. It is a hope.

Final Take: The Price of a Token Is Not the Price of Intelligence

Token prices are falling faster than almost any cost curve in technology. That is real, and it is good. It is why AI capability that was a research demo three years ago now costs fractions of a cent.

But cheaper tokens do not make AI cheap.

They make AI everywhere.

Every price drop moves more work above the break-even line. Every capable model invites bigger tasks. Every agent that can run longer will be asked to. That is how Jevons' coal story played out in the nineteenth century, and it is how tokenomics is playing out now.

The organizations that handle this well will not be the ones with the best vendor discount. They will be the ones that know what each token buys, decide deliberately what is worth automating, and enforce those decisions in their systems.

Vendors will keep making tokens cheaper.

Deciding how many you need is still your job.

Sources and Further Reading

  1. Fortune, Tokens are getting cheaper, but companies are spending even more on AI as a result, top economist warns (June 17, 2026): https://fortune.com/2026/06/17/why-is-ai-spending-increasing-as-tokens-get-cheaper-jevons-paradox/
  2. Bain & Company, How token economics will change opex: https://www.bain.com/insights/how-token-economics-will-change-opex/
  3. Epoch AI, The plunging price of thought (September 2026): https://epoch.ai/publications/the-plunging-price-of-thought
  4. Gundlach, Lynch, Mertens, and Thompson, The Price of Progress: Price Performance and the Future of AI (MIT; November 2025, revised March 2026): https://arxiv.org/abs/2511.23455
  5. a16z, Welcome to LLMflation: LLM inference cost is going down fast (November 12, 2024): https://a16z.com/llmflation-llm-inference-cost/
  6. Google, Q2 2026 earnings call: Remarks from our CEO (July 22, 2026): https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q2-2026/
  7. Google, Q3 2026 earnings call: Remarks from our CEO (October 28, 2026): https://blog.google/company-news/inside-google/message-ceo/alphabet-earnings-q3-2026/
  8. Digit, Google I/O 2026: Sundar Pichai gave numbers to explain rapid AI adoption (May 2026): https://www.digit.in/features/general/google-io-2026-sundar-pichai-gave-numbers-to-explain-rapid-ai-adoption.html
  9. Anthropic Engineering, How we built our multi-agent research system (June 13, 2025): https://www.anthropic.com/engineering/multi-agent-research-system
  10. NVIDIA, Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment: https://blogs.nvidia.com/blog/productive-durable-fungible-ai-factories/
  11. Kapoor et al., Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation (October 2025; ICLR 2026): https://arxiv.org/abs/2510.11977
  12. William Stanley Jevons, The Coal Question (1865).

About the Author

Suraj Khaitan - Gen AI Architect | Building scalable platforms and secure cloud-native systems

Connect on LinkedIn | Follow for more engineering and architecture write-ups

Has your AI spend gone up even as token prices fell? What finally helped you get it under control: budgets, routing, caching, or something else?

๐Ÿ“ฐ Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ€” full credit and traffic to the original publisher.