LLMflation

The Bill That Changed the Conversation

In May 2026, the story that changed how enterprise AI is discussed did not make every front page.

An unnamed company had reportedly run up a $500 million bill on Anthropic’s Claude in a single month. Not through negligence, exactly. Through deployment. They had given their entire workforce access to the model, with no usage caps and no governance. Agentic workflows compounded the problem. These are automated sequences where the AI plans, acts, reads the result, and acts again, re-sending the full conversation history at every step. What looked like a modest per-query cost had turned into a figure that exceeded most companies’ entire annual technology budget (Tom’s Hardware / Fast Company, May 2026).

Microsoft’s story broke around the same time and was more specific. The company had invited thousands of engineers in its Experiences and Devices division to use Claude Code starting in December 2025. By May 2026, the tool had become the preferred choice inside the division. Token costs per engineer began outpacing what Microsoft paid those engineers in salary. The company set a June 30 cutoff and redirected engineers to Copilot CLI (The Next Web, 2026).

The gap between that picture and the figures that follow is not a contradiction. An engineer chatting with a model spends like a subscriber; an engineer running agentic sessions that work for hours unattended spends like a data centre. The bill tracks the workflow, not the headcount.

Uber had got there first. Its operations chief told The Information in April that the company had burned through its entire 2026 AI coding budget in four months: roughly 5,000 engineers on Claude Code, the heaviest users spending $500 to $2,000 each per month individually (Tom’s Hardware, 2026).

None of these are outliers. They are case studies in a structural problem: the price of AI has collapsed, so every new use case looks affordable. Then you do the arithmetic on volume.

The term for the dynamic: LLMflation, coined by a16z (Andreessen Horowitz, a major Silicon Valley venture fund) in 2024 to describe the paradox where AI costs per call keep falling while AI bills keep rising. A model good enough to pass a standard knowledge test cost roughly $60 per million tokens (each token is roughly three-quarters of a word) in late 2021. By 2024 it was $0.06, a fall of roughly 1,000× in three years, faster than the PC era’s decline in compute. But AI charges per action, not per seat. Lower prices triggered far more actions, not lower bills. Token prices fell roughly 1,000× over three years (a16z, 2024) while total enterprise AI spend rose more than 300% in two (Ramp, 2026).

The Meter Never Stops

Three things are happening simultaneously inside most enterprises. They look unrelated. They are the same problem.

The first: usage consistently outruns the forecast. In 2025, 85% of companies missed their AI cost projections by more than 10%; 24% missed by more than 50% (Mavvrik / Benchmarkit, 2025). Two years ago, 31% of finance teams actively managed AI spend. Today it is 98%. The number tripled not because companies got disciplined, but because the bills arrived and demanded attention (State of FinOps, 2026).

The second: agents turn the meter into a fire hose. A standard AI chat interaction uses a few thousand tokens. A million tokens is roughly a 750,000-word document (the equivalent of asking the model to read and write War and Peace from scratch, for every complex task). An agentic workflow re-sends the entire conversation context (the AI’s working memory) at every step. Industry data puts agent token use at 10-100× a standard chat. Stanford’s Digital Economy Lab measured the most complex agentic tasks at 1,000× a standard reasoning call. Simple tool-calling agents consume 5,000-15,000 tokens per task; complex multi-agent systems consume 200,000 to over 1,000,000 tokens per task (Stanford Digital Economy Lab, 2025). Goldman Sachs estimates agents will multiply total enterprise token demand 24× by 2030.

The third: the most expensive model is rarely the right model. The price gap between the cheapest production AI model and the most expensive frontier reasoning model is currently about 4,500×. The cheapest runs at around $0.04 per million tokens. The most expensive frontier reasoning models run at $180 per million tokens. Most enterprises have deployed the top tier to everyone. Analysis consistently shows that roughly 85% of enterprise queries could be handled by budget-tier models (the ones at the $0.04 end) with no meaningful quality loss (FinOps Foundation, the industry body tracking AI infrastructure costs, 2026).

The hidden overhead compounds everything. OpsLyft (an AI cost analytics firm) found that true AI bills run 40-60% higher than the visible token line, and in the heaviest pipelines a call that looks like $0.05 can land at $0.20 once embeddings (text to searchable numbers), retrieval, re-ranking, output validation, retries, and prompt overflow are counted (OpsLyft, 2026).

This Comes Straight Off Your Margin

Inference is not a technology line item. It is cost of goods sold: the costs that come directly out of gross margin, the number investors use to price companies.

84% of companies reported that AI was already eroding gross margins by more than 6% (Mavvrik / Benchmarkit, 2025). Bessemer’s portfolio data puts AI-native gross margins at 50-60%, against 70-90% for mature SaaS, and software economics is quietly turning into manufacturing economics (Bessemer, 2025). Even OpenAI, the company whose products power much of this spending, ran $13B in revenue against $22B in costs in 2025, a $9B net loss at scale (Fortune, 2025).

The pattern is consistent: the more AI gets used, the more it costs, and the less of that cost is recovered in measurable output. This is not a price problem. It is a volume and governance problem. The companies that have protected their margins are not the ones with cheaper models. They are the ones that controlled who was using which model for what. There is a reason for that. It comes before usage.

The Access Problem

The piece most AI cost discussions miss is the one that makes every other problem harder to fix.

In 2026, Meta created internal leaderboards ranking employees by token consumption, a practice that quickly became known as “tokenmaxxing” (spending tokens competitively, not productively). Employees competed to spend the most tokens. They threw entire projects at agentic AI with no discipline about whether the output was good, whether a cheaper model would have worked, or whether the task suited AI at all. The leaderboard rewarded consumption, not outcomes (industry reporting, 2026).

Microsoft’s experience tells a version of the same story. The engineers who used Claude Code most heavily were not necessarily the ones producing the best software. Token costs outpaced salaries. And Faros AI found that code churn (lines of code deleted versus added, a measure of rework) increased by more than 800% in teams with high AI adoption. The teams spending the most on AI were producing the most waste (Faros AI, 2026).

The data points to something the economics alone do not explain: AI amplifies what you bring to it.

It is worth saying plainly that broad access to powerful tools has genuinely produced unexpected value. The marketer with no technical background who caught a product flaw through an unusual prompt. The junior analyst who used AI to try an approach a veteran would have ruled out on instinct, and was right. Ungated access is not without merit. The argument here is not for restriction: it is for matching model tier to demonstrated task competence, with a clear path to qualification for anyone willing to earn it.

A senior engineer with fifteen years of experience using Claude Code produces faster, better-reviewed code at lower cost, because they can evaluate the output, catch the errors the AI cannot catch, and prompt with precision. Evaluation means spotting when the model has hallucinated a citation (invented a source that does not exist), reversed a logical condition in code, or produced a plausible-looking answer that is wrong in a way only domain knowledge reveals. A junior developer with six months of experience using the same tool produces faster, worse code. They cannot tell good output from plausible-looking output. They iterate blindly on prompts they cannot evaluate, and burn tokens on approaches an experienced engineer would immediately dismiss.

The 4,500× price gap between the cheapest and most expensive models is not the real cost driver. The real driver is the gap between what a skilled user produces per dollar and what an unskilled user produces per dollar. A budget model in the hands of someone with deep domain expertise costs almost nothing and produces real value. A frontier reasoning model in the hands of someone without the background to evaluate its output costs $180 per million tokens and produces fast, confident, unreviewed work that someone else will eventually have to fix.

This is not a controversial idea when applied anywhere else. We do not give junior surgeons operating theatres without supervision. We do not give junior analysts sole authority over financial models that go to a board. We understand, in every other context, that tools amplify skill. Amplifying the absence of skill at speed is not progress. We have simply forgotten to apply this to AI.

Access to frontier AI models (most powerful, most expensive) is a professional resource. It should be treated like one.

Govern the Usage and the Access

The companies that have escaped LLMflation are not the ones with the cheapest models. McKinsey puts them at roughly 6% of enterprises and labels them “AI high performers.” What the published case studies of those companies consistently show is two governance layers running simultaneously: usage controls and access controls.

Usage governance: four controls that compound

Route every task to the cheapest capable model. A cascade architecture (routing tasks to cheaper models) sends simple calls to the budget tier, escalates complex ones to frontier models only when the task requires it. Analysis consistently shows this cuts inference (the cost of running the model) cost 40-70%, because most queries never need to leave the budget tier (FinOps Foundation, 2025).

Cache the answers you have already paid for. Semantic caching (reusing stored AI answers) removes about 31% of repeat queries before they reach the model. Prompt caching (reusing unchanged prompt sections) makes cached tokens roughly one-tenth the standard price. Add batch pricing on top (roughly 50% discount) and a call that would have run at $0.20 runs at about $0.01 (about 5% of standard cost) (OpenAI, 2026).

Cap the spend before it surprises you. Token budgets as hard ceilings per team and per workflow. Fix retry logic that can loop indefinitely. Define maximum context windows for each use case. The $500M bill happened because there were no ceilings.

Measure cost per outcome, not cost per call. “Cost per call” tells you almost nothing. “Cost per resolved customer query,” “cost per code review,” “cost per draft approved”: these are the numbers that tell you whether the spend is working. The vendors that price this way (Intercom at $0.99 per resolved query, Salesforce Agentforce at roughly $2 per conversation) already know their cost per outcome. The discipline turns the meter into a dial (Intercom, 2026).

Access governance: the layer most companies skip

Tier model access by role and demonstrated competence. Not everyone needs frontier reasoning. Not everyone should have it. The qualification for moving up a tier is not seniority: it is demonstrated domain knowledge. The ability to evaluate what the model produces in that domain and catch what it cannot catch.

The upgrade is available to anyone. A junior developer on the budget tier earns frontier access when they can demonstrate they can evaluate what a frontier model produces in their domain: spotting the errors, assessing the quality, catching the hallucinations. That is not hierarchy preservation. It is competence development with a measurable gate.

The advantage is real, and it is unequal: when skilled engineers have frontier access and unskilled ones are routed to budget tiers, the skilled engineers work faster and the budget recovers its balance. The companies with ungated access to everything get neither the speed nor the savings.

The cost reduction from running both governance layers together reaches 60 to 90 per cent against ungoverned baseline (FinOps Foundation, 2026). Most companies are still paying full price. Not all of them think that is the wrong call.

Bubble, or Breakthrough?

Three honest positions, all with something right in them.

The Price of the Wrong Hands