AI Token Prices Plunge 90%, Yet Enterprise Bills Double: The Jevons Paradox Is Repeating

A Paradox That Confounds CTOs

The price of a single AI token has fallen by more than 90% since 2023, yet over the same period, total enterprise spending on large language models has doubled. These figures come from Apollo chief economist Torsten Slok, based on the Silicon Data Token Expenditure Index. Payment platform Ramp's billing data offers even more direct evidence: enterprise monthly AI spending roughly quadrupled between February 2025 and February 2026.

This is not a paradox on the horizon—it is a paradox that has already arrived. If teams still treat "cost per million tokens" as their core budgeting metric, they are optimizing the wrong variable from the very start.

Explaining the Jevons Paradox Through Your Bill

In 1865, economist William Stanley Jevons observed a counterintuitive phenomenon: after steam engine efficiency improved and the real cost of coal fell, Britain did not burn less coal—it burned more. Efficiency gains pushed coal-powered industry into scenarios that had previously been too costly, and total consumption exploded.

In his June 2026 Daily Spark, Slok described the current state of AI spending as "the Jevons Paradox playing out." The mechanism is identical: once tokens become cheaper, companies don't spend less—they run more AI agents, automate more workflows, generate more code, and attempt tasks that were previously shelved as too expensive.

The formula is almost brutally simple:

Cost per task = tokens per task × cost per token

Cost per token is falling, but tokens per task is rising far faster—and that is the variable no one factors into the budget.

A 2026 analysis by Bain & Company found that from December 2024 to December 2025, token costs fell by half, but token consumption grew by 450% over the same period. Bain attributed this to three factors: companies upgraded to newer models that are more capable but also more token-hungry; agents took on more complex multi-step tasks; and teams expanded usage once they trusted AI to handle larger jobs. Bain's conclusion was blunt: "The bills remain stubbornly high."

Bain's core finding: token costs halved in a year, yet token consumption grew 450%. The bill didn't disappear—it shifted from "paying more per unit" to "paying for far more units."

The Jevons Paradox, also known as the Jevons effect, was proposed by British economist William Stanley Jevons in his 1865 book The Coal Question. He observed that after Scottish steam engines were improved and efficiency rose sharply, the coal required per unit of output fell, yet Britain's national coal consumption rose rather than fell. The reason: improved efficiency lowered the real cost of using coal, allowing it to enter application scenarios (such as metallurgy, railways, and power generation) that had previously been excluded as too expensive, and the demand added by these new scenarios far exceeded the savings from any single scenario. This paradox was later widely applied in energy economics (e.g., global lighting electricity use rose after LED lighting became widespread) and in information technology (e.g., falling cloud computing unit prices spawned larger-scale cloud spending). In the AI context, its core insight is: falling unit costs and rising total spending can coexist for a long time, so any budget model centered on unit price will systematically underestimate actual spending. Understanding this mechanism is a prerequisite for judging the trajectory of AI bills.

What the Community Is Discussing

The "per-token trap" was one of the most heated topics on Hacker News in 2026. After Fortune reported on the Uber budget incident, the related HN thread drew hundreds of comments, many from engineers who had personally experienced cost shocks.

The recurring theme: engineers know that per-token pricing is misleading, yet their organizations still use it as a primary metric. One commenter pointed out that "the vast majority of the cost is dominated by the cost of cached input tokens in Claude Code sessions"—what really drives up expenses is repeated context reads, not model choice itself.

Another post titled "Tokenmaxxing is dead, long live tokenmaxxing" captured the industry's turning point: companies were taking down internal usage leaderboards and shifting to outcome-oriented measurement.

Meanwhile, threads on "AI sticker shock" and the "AI affordability crisis" show that the practitioner community is increasingly skeptical of the "AI will eventually be free" narrative and is demanding better cost tools.

How Agents Break the Budget

The shift from chatbots to agents is the single biggest driver of the cost explosion. A chatbot query is a one-off interaction: send a prompt, get a reply, pay for a few hundred tokens. But AI agents now embedded in coding tools, customer service pipelines, and enterprise automation operate in a completely different way.

In an August 2026 press release, Gartner coined the term "Inference Paradox." Its forecast: even as per-token prices keep falling, the inference cost per agent workflow will grow more than fivefold by 2028.

The reason is that agents resend the full context at every step. Each tool call, each reasoning iteration, each self-correction loop multiplies the token count. Analyst Koo Ping Shung laid out the arithmetic: a simple AI query uses a few hundred tokens; multi-step reasoning and agentic coding tasks use 5 to 30 times more; and for the most complex autonomous workflows—the kind Uber deployed to thousands of engineers—consumption can reach 1000 times the baseline.

At a flat fee of $30 per seat, the math still works. But under usage-based pricing, a single user's agent session can trigger hundreds of dollars in compute costs. And in 2026, every major AI provider's most powerful models have shifted to usage-based billing.

Gartner's Inference Paradox: token costs will fall 95% by 2030, while agent workflow costs will rise more than 5x by 2028. Both statements are true at the same time—that is exactly where the trap lies.

Prompt Caching is an optimization mechanism offered by mainstream large model APIs: when multiple requests share the same prefix content (such as system prompts, long documents, or codebase context), the provider can cache the computation results for that portion and reuse them in subsequent requests, avoiding redundant computation. For scenarios billed by input tokens, cached input tokens are typically charged at 10% to 25% of the original price, and some vendors offer even lower tiered pricing for cache writes/reads. This mechanism is especially critical for agents, because agents resend the full context at every tool call and self-correction loop, and uncached repeated input is often the main component of the bill—as the HN commenter in this article noted, the cost of Claude Code sessions is dominated by cached input tokens. Designing prompts well (putting stable content in the prefix and variable content at the end) can significantly improve cache hit rates and is one of the most direct cost levers for engineering teams.

Uber's Cautionary Tale

No company illustrates the per-token trap better than Uber.

In early 2026, Uber rolled out Claude Code to thousands of engineers. They built an internal leaderboard that ranked employees by token consumption—a practice the industry now calls "tokenmaxxing." The result: Uber burned through its entire annual AI coding budget in the first four months of 2026.

In May, Uber COO Andrew Macdonald admitted that the link between AI spending and shipped product features "has not yet been established." CTO Praveen Neppalli Naga eventually declared the end of the tokenmaxxing era, telling Fortune: "AI success won't be measured by who burns the most tokens."

But how Uber recovered is more instructive. Rather than cutting AI usage, Naga's team treated the spending problem as an engineering problem:

1. Improved prompt caching—reducing redundant context re-reads
2. Changed default models—routing routine tasks to cheaper models
3. Hourly usage visibility—letting engineers see their own costs in real time
4. Model efficiency testing—evaluating new models not just on capability but on token economics

The result: Uber lowered its cost per token while quadrupling the number of employees using frontier AI tools. This is the Jevons Paradox working for you rather than against you—provided you measure it correctly.

FinOps (Cloud Financial Operations) is a set of cost management practices spanning finance, engineering, and business teams. It originally emerged to address runaway public cloud spending and emphasizes observability, budget attribution, and unit economics metrics to govern usage-based IT resources. As AI workloads have proliferated, the FinOps Foundation has incorporated AI cost management into its framework, with core practices including: tagging costs by team or product line, building real-time usage dashboards, setting budget thresholds and alerts, and measuring efficiency by "unit cost" (such as per successful task or per thousand inferences) rather than by total volume. The article mentions that 73% of organizations exceed their AI cost budgets and that the share of FinOps teams managing AI spending rose from 31% to 98% in two years—a reflection of this practice's rapid expansion in the AI era. Understanding FinOps helps grasp the organizational and institutional foundation behind the metric shift "from cost per token to cost per successful outcome."

Ramp Data: 4x Growth in 12 Months

Corporate card and expense management company Ramp sits on a unique dataset: real billing records from tens of thousands of businesses. Its April 2026 Leading Indicators report revealed the spending reality beneath the headline price declines:

• On the Ramp platform, adoption among companies paying for AI surpassed 50.4% in March 2026, up from 35% a year earlier
• Monthly AI spending quadrupled between February 2025 and February 2026
• For the highest-spending companies, AI costs jumped 50% or more in roughly one out of every four months
• The number of enterprise customers spending over $1 million annually with Anthropic has doubled to 1,000 since February 2026

This is first-party billing data from actual corporate credit cards, not surveys or forecasts. When more than 50% of businesses are paying for AI and spending is growing fourfold year over year, falling per-token prices are merely a rounding error for real bills.

Ramp's earlier data also showed that companies investing heavily in AI are growing revenue faster than those that aren't—though they note that correlation is not causation. That is precisely what makes the trap dangerous: the spending feels productive, making it harder to question.

The counterargument: is the "AI cost explosion" exaggerated?

The cost explosion narrative benefits specific groups: FinOps vendors, AI cost management startups, and consultancies selling optimization services. For the median company, AI spending is still a small fraction of total IT spending. The companies making headlines—Uber, Meta, and large tech firms with thousands of engineers—are outliers that gamified token usage with leaderboards and "tokenmaxxing" culture. This is a management failure, not a structural economic problem. The real question is whether the value created justifies the spending, and on that question, the data is murkier than the cost doomsayers admit.

The Formula That Actually Matters

The industry is slowly converging on a solution, and it begins with changing the metric. Slok, Gartner, and Uber's postmortem all point to the same shift:

Stop measuring cost per token. Start measuring cost per successful outcome.

1. Model Routing, Not Model Loyalty

Not every task needs a frontier model. A classification task that GPT-4.1 Nano can handle at $0.10 per million tokens shouldn't be routed to Claude Opus 5 at $15 per million tokens. Bain found that teams upgrade to the most expensive model the moment a new one launches—even when a cheaper model performs identically on their use case.

Build a routing layer: send hard tasks to frontier models, routine tasks to efficient models, and use cached responses for content you've seen before.

2. Set Token Budgets by Team

Give each team a monthly token budget and real-time visibility. When engineers can see that their last agent session cost $47, behavior changes quickly. That's exactly where Uber's turnaround began.

3. Caching and Context Management

Agents resend the full context on every turn. Prompt caching, summarization of historical context, and smart window management can cut token volume by 30% to 60% without sacrificing task quality. This is the single highest-ROI optimization engineering teams can make.

4. Measure Output, Not Usage

Uber built a leaderboard around token consumption and got exactly what it incentivized: maximum consumption. The fix is to flip the metric from "tokens used" to "tasks delivered per dollar of token spend." The companies that win in the post-tokenmaxxing era will treat AI spending as an input to the productivity equation, not a scoreboard.

What This Means for You

If you're an engineering leader, CTO, or CFO watching your AI bills, here's the bottom line:

Per-token prices will keep falling. Goldman Sachs forecasts token consumption will grow 24x by 2030, while Gartner expects per-token costs to fall 95% over the same period. But agent workflows—the only AI deployment model that truly replaces human labor—will consume tokens faster than any price decline.

This is not a temporary mismatch; it's the new equilibrium. The AI industry has entered a Jevons cycle, much like electricity, telecommunications, and cloud computing before it: units get cheaper, usage explodes, and total spending grows.

The winners won't be the companies that spend the least on AI, but those that spend just right—replacing tokenmaxxing-style permissiveness with per-team budgets, model routing, caching, and outcome-oriented metrics.

As Uber's CTO put it: efficiency is an engineering problem, not a budget problem. Solve it the way you'd solve an engineering problem.

The one-line takeaway: track cost per successful outcome, not cost per million tokens. Every other optimization starts with getting this metric right.

One-line Summary

Per-token prices have plunged 90%, yet enterprise AI bills have doubled—the Jevons Paradox is repeating.

Paragraph Summary

The article centers on a paradox that has already materialized: the price of a single AI token has fallen more than 90% since 2023, yet total enterprise spending on large models has doubled over the same period, and Ramp data shows monthly AI spending quadrupled within a year. The author uses the Jevons Paradox to explain this—efficiency gains let AI enter more scenarios that were previously too costly, and total consumption explodes accordingly. Agents are the biggest driver, since they resend the full context at every step, and complex workflows can consume up to 1000x the baseline, prompting Gartner to coin the "Inference Paradox." Uber's case is the most cautionary: an internal token consumption leaderboard led to burning a full year's budget in four months, after which prompt caching, model routing, real-time visibility, and efficiency testing cut per-token costs while quadrupling usage. The article ultimately argues for abandoning the per-token cost metric in favor of cost per successful outcome.