Token Prices Fell 98%. Developer AI Bills Rose 100×. The Jevons Paradox Has Arrived.
Microsoft killed its Claude Code licenses after burning through an annual AI budget in months. Uber blew its 2026 AI spend in 120 days. Gartner says AI coding costs will surpass developer salaries by 2028. I ran the numbers on how we got here.
Roughly five thousand Microsoft engineers loved Claude Code so much they burned through their division's entire annual AI tools budget before spring ended. Per-engineer API costs hit $500 to $2,000 per month, and Rajesh Jha, the EVP of Experiences and Devices, sent the memo every CTO fears writing: kill the licenses, switch to something cheaper, deadline June 30. At the same company, GitHub had just switched Copilot to usage-based billing on June 1 because flat-rate subscriptions were hemorrhaging money from heavy users. A tool built to make developers more productive had become too expensive for the company that makes the tool.
Microsoft is not an outlier; it is a bellwether.
The Budget-Burn Cascade
In April, Uber CTO Praveen Neppalli Naga confirmed to The Information that the ride-hailing company had exhausted its entire 2026 AI coding tools budget in four months. Its irony was pointed: Uber had been running internal leaderboards ranking engineering teams by their AI tool consumption, gamifying the very spending that blew the budget. At Nvidia, Bryan Catanzaro, VP of Applied Deep Learning, told Axios that for his team "the cost of compute is far beyond the costs of the employees." And one unnamed enterprise reportedly spent $500 million in a single month after failing to set usage limits on Claude licenses.
These are not cautionary tales from companies that mismanaged deployment but rather the results of the largest, most technically sophisticated organizations on Earth running exactly the playbook the AI vendors designed for them, then recoiling when the bill arrived.
The Jevons Paradox, Quantified
In 1865, William Stanley Jevons observed that James Watt's more efficient steam engine did not reduce coal consumption. It increased it, because lower costs per unit of work made coal-powered machinery economical for far more applications, which is precisely the mechanism now gutting enterprise AI budgets across the software industry.
Per-token prices have cratered. When GPT-4 launched in March 2023, output tokens cost approximately $0.06 per thousand; Claude 3.5 Sonnet, released in 2024, charged $0.015 per thousand output tokens; and by mid-2026, lightweight models like GPT-5.4 nano cost about $0.00125 per thousand output tokens through GitHub Copilot. That is a 98% decline in unit cost, the kind of deflationary curve that normally triggers an adoption boom and savings windfall.
What happened instead was an adoption boom and a spending explosion, because the nature of AI-assisted coding fundamentally changed. In 2023, Copilot was an autocomplete engine: a developer typed, the model finished the line, consuming perhaps 50,000 tokens per developer per month. By 2026, agentic coding tools run autonomously for hours, reading entire codebases, iterating through multi-file refactors, maintaining context windows approaching one million tokens per session, and spawning sub-agents for code review, testing, and documentation. Token consumption per developer has risen an estimated 100× to 1,000× from 2023 levels, from tens of thousands to tens of millions of tokens per month.
The math is brutal: tokens are 48× cheaper, but developers consume 100× to 1,000× more of them. Net spend per developer rises 2× to 20× despite the price collapse. William Jevons is nodding from the grave.
The Crossover Calculation
On June 24, Gartner published a prediction that AI coding costs per developer will surpass the average developer's salary by 2028, which sounds outlandish until you see the trajectory laid out in publicly available data.
| Year | Dominant Tool | Monthly Cost/Dev | Annual Cost/Dev | % of Avg. Salary ($83K) |
|---|---|---|---|---|
| 2023 | Copilot (flat seat) | $19 | $228 | 0.3% |
| 2024 | Copilot + Claude chat | $19–$39 | $228–$468 | 0.3–0.6% |
| 2025 | Claude Code, early agentic | $100–$500 | $1,200–$6,000 | 1.4–7.2% |
| 2026 | Full agentic (MSFT actuals) | $500–$2,000 | $6,000–$24,000 | 7.2–28.9% |
At Microsoft's reported midpoint of $1,250 per month ($15,000 per year), AI coding tools already consume 18% of the average US developer salary of $83,201 according to PayScale. At the high end, $2,000 per month reaches 29%. That's before the trajectory bends upward.
Projecting forward with a conservative 2× annual growth rate in per-developer AI spend (well below the 4× seen from 2024 to 2026), the crossover with the national average developer salary arrives in 2029. If growth follows the steeper historical 4× trajectory, tool cost exceeds salary in 2027, overshooting even Gartner's aggressive timeline. Between late 2027 and early 2029, the landing zone narrows depending on how successfully enterprises implement governance.
Bay Area math is even more unfavorable. San Francisco's median total developer compensation sits at $274,500 per Levels.fyi as of June 28, 2026. But this salary premium is exactly what has driven the most intensive AI adoption: the $274,500 developer is the developer whose employer gave them unlimited Claude access, ran the leaderboards, and then panicked when the bill came.
The June 1 Shock
GitHub's June 1 transition to usage-based billing provided a real-time controlled experiment in cost visibility. Under the old system, developers consumed "premium requests" without seeing token counts, a flat-rate illusion that hid the actual compute cost of their work. Under the new system, one AI Credit equals one cent, and every interaction has a visible price.
Results were immediate. TechSpot reported that developers on the $10/month Pro plan (which includes 1,500 credits, or $15 of usage) were watching credits evaporate within hours. One user burned 21% of their monthly allocation on day one. Gone. Another spent 840 credits on a single cautious session with Claude Sonnet 4.6, and a pair of Copilot-driven commits consumed 5,000 credits, a quarter of the top-tier Max plan's entire monthly budget of 20,000, which means two code changes cost a developer the equivalent of four days' worth of tool access.
Model choice matters enormously, and not in the way enterprise buyers assumed when they signed usage-based contracts. One million output tokens from GPT-5.4 nano costs about $1.25 through Copilot; the same volume from GPT-5.5 costs approximately $30, a 24× difference for the same number of tokens. But developers don't choose models the way they choose rental cars. They pick the model that produces the best code, and frontier models produce observably better code, so consumption gravitates toward the expensive end of the menu precisely when the billing model starts punishing that behavior.
The Strongest Case Against Panic
The bulls have a genuine argument, and it would be dishonest to dismiss it. Token prices are still falling, inference efficiency is improving with every hardware generation, and cheaper models are narrowing the quality gap with frontier offerings. One developer on TechSpot's comment thread reported integrating DeepSeek into their workflow and estimating costs at "about 7 cents for 15 million tokens," a figure that makes GitHub Copilot's pricing look almost predatory by comparison. Companies can route tasks to appropriate model tiers. Governance frameworks, which Gartner itself recommends, can cap runaway spending. Microsoft's Claude Code debacle may be a growing pain, not a structural constant: the company pushed hard, didn't monitor spending, and got burned by a pricing model transition. That's a management failure, not a law of physics.
Jensen Huang has framed the problem entirely differently, arguing that AI token budgets should function like a capital allocation: give every elite engineer a token budget worth half their salary and measure the output multiplier. If a $400,000 engineer produces $1.2 million in value using $200,000 in tokens, the arithmetic works even though the tool costs more than a junior hire. The question isn't whether tokens are expensive; it's whether they produce more value than the next-best use of that money. Nobody complains that a CNC machine costs more than the machinist.
What This Analysis Doesn't Cover
Microsoft's $500-to-$2,000-per-month range comes from secondhand reporting by The Verge and industry analysts, not from a Microsoft 10-K filing, and the company has not publicly disclosed per-engineer API costs. Uber confirmed the budget exhaustion but not the budget amount, and the $500-million single-month figure comes from an anonymous source to Axios whose identity and client cannot be independently verified. My token consumption estimates for agentic workflows rely on ranges (5× to 30× more tokens than chat), not precise telemetry, and the Jevons Paradox framing assumes consumption growth persists at something like recent rates when governance interventions could break the pattern. Different salary benchmarks produce different crossover dates: PayScale's $83,201 average puts the crossover closer; Levels.fyi's $274,500 SF median pushes it further out.
What You Can Do About It
If you manage an engineering budget, install token monitoring before you install AI coding tools, not after, because Microsoft and Uber learned the hard way that adoption without observability is a blank check drawn on a corporate account with no spending cap. GitHub's new credit system at least makes costs visible; use it as a dashboard, not a surprise. Set per-developer monthly ceilings and route routine tasks like boilerplate generation, documentation, and test scaffolding to cheaper models while reserving frontier models for complex architecture work where the quality differential justifies the 24× price premium.
If you are a developer, understand that your token consumption now appears on a ledger somewhere, and your employer is reading it with the same attention they once reserved for cloud compute bills and corporate travel expenses. The era of "let Claude think about it for a while" with a million-token context window is ending for the same reason the era of unlimited data plans ended: the economics don't work when power users consume 100× what casual users do. Learn context engineering. Smaller, more precise prompts produce better results and cost less, and killing long-running chat threads that accumulate massive input token histories is no longer a nice-to-have optimization but a career-preserving habit.
If you are evaluating AI coding tools for a new deployment, benchmark against total cost of ownership, not subscription price, because a $39/month Copilot Pro+ subscription is a rounding error while the credits that heavy agentic usage consumes on top of it are not. Budget at $500/month per developer as a baseline for teams running agentic workflows, and stress-test the estimate upward before you commit.
The Jevons Paradox has no villain. Cheaper tokens and more powerful coding agents are both genuinely good. But the interaction between falling unit prices and exploding unit consumption produces a result that surprises everyone who only looked at one half of the equation, and the enterprise CFOs discovering this asymmetry in their quarterly budget reviews are learning the same lesson that Victorian-era economists learned about coal: making a resource cheaper does not make it cheaper to use, because the uses multiply faster than the prices fall.