How Much Companies Spend on AI Tokens in 2026

In December 2025, Uber rolled out Claude Code to its engineering team. By February, usage had doubled. By April, the ride-hailing company had burned through its entire 2026 AI budget. Adoption among Uber’s 5,000 engineers jumped from 32% to 84% in four months, and the monthly AI token bills followed: $500 to $2,000 per engineer, with no ceiling in sight. COO Andrew Macdonald told investors he couldn’t draw a line between the token spending and consumer-facing results. “That link is not there yet,” he said.

Uber isn’t alone. Microsoft pulled Claude Code licenses from its Experiences & Devices division after costs hit roughly $2,000 per engineer per month. One unnamed enterprise reportedly racked up $500 million in a single month on Claude after failing to activate spending caps. The pattern is the same everywhere: companies went all-in on AI tokens in 2025, and 2026 is when the bill arrived.

AI token spending is the total cost companies pay to process inputs and outputs through large language models like Claude, GPT-4, and Gemini, measured in tokens (roughly 0.75 words each). In 2026, it’s become one of the fastest-growing line items on corporate balance sheets.

Last updated: June 2026

Quick answers

How much do companies spend on AI tokens?

The median company spends $11.38 per employee per month on AI tokens, according to Ramp’s June 2026 AI Index, which tracks spending across 70,000+ U.S. businesses. The top 1% of firms spends $7,450 per employee per month. Overall median monthly company spend is $2,246, but the average is $140,842, because a small number of heavy users pull the average far above the median.

Why are AI costs going up if tokens are getting cheaper?

Token prices dropped over 90% since 2023, but companies respond by using more AI, not less. Apollo chief economist Torsten Slok calls this Jevons paradox in action: as the unit cost of intelligence collapses, companies run more agents, automate more workflows, and generate more code, pushing total spending higher even as each token costs less.

How much do companies spend on AI tokens per employee?

The spending gap is staggering. Ramp’s June 2026 AI Index, drawn from transaction data across more than 70,000 U.S. businesses, reveals a 680x difference between the most AI-aggressive firms and everyone else.

The top 1% of companies spends $7,450 per employee per month on AI. The top 10% spends $611. The median firm spends $11.38, roughly the cost of a single ChatGPT or Claude Enterprise seat. For the vast majority of American businesses, AI adoption still means one subscription per employee and not much else.

That said, the top 1% is accelerating. Their per-employee AI spend grew 14.1% in a single month, according to Ramp economist Ara Kharazian’s June report. Even at $7,450 per head, though, no company is spending more on AI than on engineering salaries. The “spend as much on AI as you do on a developer” talking point from early 2026 hasn’t materialized in the data.

The spending also breaks down differently by company size. For most small and mid-size businesses, the majority of AI costs go to SaaS subscriptions: ChatGPT Enterprise, Claude for Business, Copilot licenses. These are predictable monthly costs. For larger companies and AI-native startups, the majority flows through API billing, where costs scale with usage and can spike unpredictably. That distinction matters for budgeting. A 50-person startup paying $20/seat/month for ChatGPT Enterprise knows its bill before the month starts. A 50-person AI-native company routing thousands of API calls through Claude and GPT-4 might see bills swing 40% month to month based on project load.

For founders building AI wrapper startups or running lean teams, the relevant benchmark is the top 10% figure: $611/employee/month. That’s the zone where teams are using multiple frontier models, running agentic workflows, and treating AI as core infrastructure rather than a productivity perk. If your company is spending less than $50 per employee, you’re in the bottom half of adoption.

Why are AI token bills rising when prices are dropping?

Token prices have cratered. Per-token costs for frontier models fell over 90% between 2023 and mid-2026. Claude Sonnet 4.6 costs about $3 per million input tokens. GPT-4 equivalents have followed similar trajectories. Yet corporate AI spending doubled since late 2025, per the Silicon Data Token Expenditure Index.

The explanation is a 160-year-old economic principle. In 1865, William Stanley Jevons observed that when the Watt steam engine made coal more efficient, coal consumption didn’t fall. It skyrocketed, because cheaper energy opened up uses that weren’t economical before. Apollo chief economist Torsten Slok applied this directly to AI tokens in a June 2026 Fortune analysis: “As tokens get cheaper, companies don’t spend less but instead run more AI agents, automate more workflows, and generate more code.”

Three specific forces are driving this:

First, agentic AI consumes tokens at a rate that dwarfs traditional chatbot usage. A single agentic coding session can push 400,000 to 2 million tokens through the API. Per-developer AI consumption rose about 18.6x in nine months, largely because of tools like Claude Code, Cursor, and Codex that run multi-step workflows autonomously. A chatbot answers once. An agent retries, verifies, iterates, and runs around the clock.

Second, companies upgrade to more capable (and more expensive) models rather than pocketing savings from price drops. When a better model ships, teams switch to it instead of staying on the cheaper option, because the ROI comes from capability, not cost savings.

Third, tokens per query have increased. Complex tasks generate longer prompts and longer responses. As companies ask AI to handle harder, higher-stakes work, each request consumes more raw compute.

The raw numbers confirm how quickly these forces compound. Token consumption across Ramp’s customer base grew 1,001% from January 2025 to April 2026, a 10x explosion in just 15 months. That growth concentrated in the companies already spending the most: the top quartile of AI spenders increased their token volume even faster than the broader base, widening the gap between heavy adopters and everyone else.

The result, as one TechCrunch investigation put it: “The models get cheaper. The usage gets heavier. The bill stays stubbornly high.”

What is the average AI token cost per month?

The “average” is misleading, and understanding why matters for any founder trying to benchmark their own spend. Ramp’s April 2026 data showed a median monthly company spend of $2,246 and an average of $140,842. That 63x gap exists because AI spending follows a power-law distribution: most companies spend very little, and a tiny number spend millions.

Here’s how spending breaks down across different measures, based on Ramp’s data from 70,000+ businesses:

The per-employee number is the most useful benchmark for founders. If you’re running a 10-person startup and spending $200/month total on AI, you’re at $20/employee, which puts you slightly above the U.S. median but well below the zone where AI starts replacing headcount or automating core workflows. Teams in the top 10% ($611+/employee) are the ones reporting that AI handles meaningful chunks of their engineering, support, or content operations.

For context, 40% of surveyed companies now spend more than $10 million per year on AI, per a 2026 enterprise survey. Global AI software spending is projected to reach $2.59 trillion this year, a 47% year-over-year increase. The Federal Reserve Bank of Atlanta found that per-employee spending on AI rose 50% over 2025 levels, reaching an anticipated $2,068 per employee for 2026.

How much does AI coding cost per developer?

AI coding tools are the single biggest driver of the token spending surge, and the costs vary wildly depending on how developers use them.

Anthropic’s own enterprise data puts the typical Claude Code cost at $150 to $250 per developer per month, with 90% of users under $30 on any active day. That’s the inline-completion, question-answering use case. The numbers change when developers start running agentic workflows, where the AI iterates through multi-step coding tasks autonomously. In agentic mode, a single task can consume 400,000 to 2 million tokens, and power users running multiple sessions per day reach $500 to $2,000 per month.

The premium subscription tier has converged at $200/month across the industry: Claude Code Max, Cursor Ultra, and ChatGPT Pro all hit that price point. Below that, experienced developers report spending $150 to $400/month when mixing 2-3 tools simultaneously.

Chamath Palihapitiya’s software startup, 8090, is a case study in how fast these costs compound. Since November 2025, the company’s AI costs tripled and are trending toward $10 million per year. Palihapitiya pointed to Cursor specifically as one of the biggest cost drivers, calling out “Ralph Wiggum loops” where prompts get repeatedly sent to a model until it produces a working solution. He publicly said 8090 needs to migrate off Cursor to control spending.

For founders weighing their own AI tool stack, the lesson is clear: subscription tiers give you cost predictability, but API-based token billing can spiral if agentic usage isn’t capped. Most teams that run into budget problems didn’t set spending limits before enabling access.

The companies that blew their budgets

The Uber story is the most documented, but it’s not the most extreme.

Uber rolled out Claude Code in December 2025 and actively encouraged engineers to adopt it. Usage doubled by February 2026. By April, the company had exhausted its entire annual AI budget. Engineers were generating monthly API costs between $500 and $2,000 each, across a workforce where 84% of 5,000 engineers were now using the tool. Uber responded by capping all employees at $1,500 per month per AI coding tool.

Microsoft’s Experiences & Devices division ran a Claude Code pilot starting in December 2025. By June 2026, leadership ordered engineers off the tool entirely, citing token billing that hit approximately $2,000 per engineer per month and blew through the division’s annual AI allocation.

The most extreme case, reported by Axios, involved an unnamed enterprise that spent $500 million on Claude in a single month. The company had failed to activate Anthropic’s built-in spending caps before rolling out licenses. Employees had no usage limits, and developers running long coding sessions, AI agents executing chained workflows, and employees generating large-context prompts consumed enormous volumes of tokens in surprisingly little time. The spending controls existed. They just weren’t turned on.

These aren’t edge cases anymore. OpenAI’s head of enterprise told TechCrunch that conversations with customers shifted from capability to cost control: “Six months ago, it was about capability. Now it’s spending visibility, auditability, token controls, and model efficiency.”

What disciplined teams are doing about AI token costs

The companies that aren’t blowing their budgets share a few common practices.

Per-user and per-team spending caps are the first line of defense. Uber’s $1,500/tool/employee/month cap is becoming an industry template. Anthropic’s Enterprise platform includes spend caps at the organization and user level, real-time dashboards, and role-based model access, so teams can restrict which employees use expensive frontier models versus cheaper alternatives.

Model routing is the second lever. Businesses building on AI are increasingly sending simple queries to smaller, cheaper models (including DeepSeek) and reserving frontier models for tasks that require top-tier reasoning. Ramp’s data shows firms are “increasingly opting for cheaper AI models” even as total spend rises, which is cost discipline applied at the margin rather than across the board.

Cost-per-output metrics are replacing cost-per-token as the measure that matters. Instead of tracking how many tokens a team burns, disciplined operators measure what those tokens produce: lines of code shipped, support tickets resolved, content pieces generated. If a developer burns $1,500/month in tokens but ships 3x more code, the cost per output may still be lower than the pre-AI baseline.

The Tokenomics Foundation, launched by the Linux Foundation on June 3, 2026, aims to formalize these practices into open standards. The Foundation is building canonical definitions for AI token economics, standardized metrics for usage and billing, and new measurement frameworks like cost-per-intelligence and tokens-per-watt. Early backers include Oracle, Google, Microsoft, Accenture, Booking.com, JPMorganChase, KPMG, Salesforce, SAP, and ServiceNow.

For founders running smaller teams, the practical takeaway is straightforward: set spending caps before you grant access, audit your monthly bill against what got shipped, and don’t assume cheaper tokens mean cheaper bills. If you’re building an AI consulting practice or any AI-forward business, token cost management is now a core operational skill, not an afterthought.

Where AI token spending goes from here

The spending is going up. Way up. And for anyone watching the broader AI spending debate, the trajectory is only steepening.

Goldman Sachs Research projects that global token consumption will multiply 24x between 2026 and 2030, reaching 120 quadrillion tokens per month. The primary driver is agentic AI: autonomous agents that monitor their environment continuously, call external tools, verify outputs across multiple rounds, and run 24 hours a day without human prompting. On the consumer side alone, agentic AI is expected to multiply token consumption 12x by 2030.

The saving grace for corporate budgets is that semiconductor providers are delivering cost reductions of 60% to 70% per year per token for inference. Goldman sees this creating a “gross margin inflection” for hyperscalers and model providers as revenues outpace costs. Token prices will keep falling. The question is whether usage growth outpaces the price drops, and every data point from 2026 says it does.

For founders, the calculus is simple. AI token spending isn’t a phase that passes. It’s a permanent operating cost that’s growing faster than most companies budget for. The businesses that treat it like they treat cloud infrastructure, with monitoring, caps, and cost-per-output accountability, will stay ahead. The ones that hand out unlimited API keys and hope for the best will end up like Uber in April: staring at an empty budget with eight months left in the year.