Uber spent its entire 2026 AI budget in four months, and the rest of the industry is now metering tokens

Uber gave every engineer Claude Code in December and had spent its entire 2026 AI budget by April. CTO Praveen Neppalli Naga confirmed the overrun to The Information. Nobody misused the tool. The engineers did exactly what they were asked to do, and the bill arrived anyway.

That’s the shape of the problem now hitting engineering orgs across the industry. Token pricing scales with enthusiasm, so a successful rollout and a blown budget are the same event viewed from two departments. Software has spent 20 years being a seat you buy once, whether the person logs in daily or never. AI is metered, per request, per token, per retry, and finance teams are discovering that nobody in the building can forecast it. Uber’s CTO put the structural problem plainly: “Token-based consumption pricing does not behave like the software line items chief financial officers know how to model.”

How the budget actually broke

Uber’s adoption curve is the useful part, because it’s not a story about waste. Claude Code went org-wide in December 2025. By February, 32% of engineers were on it. By March, 84% counted as agentic coding users, and by spring 95% of Uber engineers were using AI tools monthly, according to Forbes’ account of the CTO’s disclosure. Around 70% of committed code came out of AI tools. Roughly 11% of live backend updates were written by agents with no human in the loop.

The per-developer numbers look modest right up until you multiply them. Average monthly spend ran $150 to $250. Power users hit $500 to $2,000. Neppalli Naga burned $1,200 in a single two-hour demo session, which is a genuinely instructive data point: the person who owns the budget could not personally predict the cost of two hours of his own work.

Uber also did the thing that makes this funnier and worse. It ranked engineers on internal leaderboards by Claude Code usage. Every incentive in the building pointed at consuming more tokens, and the company got precisely what it measured. It has since limited employee use of tools including Claude Code and Cursor, per 404 Media’s reporting.

The consumption isn’t where you think

The comfortable story is that superhuman engineers generate mountains of code and that’s where the money goes. Accenture’s internal data says otherwise. “We’re seeing from some of the data internally at least that it’s actually not our engineers that are driving the token consumption,” Justice Kwak, who leads the firm’s agentic AI strategy, told 404 Media. The heavy users are non-technical staff doing routine document work, turning PDFs into slide decks and similar chores that a $20 tool handled fine last year.

That reframes the whole cost-control conversation. If engineers were the problem, capping engineers would fix it, and you’d trade some velocity for predictability. If the burn is a marketing analyst reformatting a deck through a frontier model, then the fix isn’t a cap at all. It’s routing, defaults and a bit of education, none of which touch the workflows actually producing value.

Vendor behavior is shifting under the same pressure. GitHub moved to per-token charging instead of a flat subscription. Microsoft is introducing budget limits while still describing itself as an AI-first company. Both moves push metering down to whoever is closest to the spend, which is the only place it can realistically live. Note what neither company did: cut the price. The seat-based era trained buyers to expect a number they could put in a spreadsheet once a year, and the vendors have decided that era is over rather than trying to reconstruct it. Every tool that used to cost a fixed amount per head is drifting toward a meter, and coding agents are simply the first category where the meter runs fast enough for anyone to notice inside a single quarter.

Two strategies, one bill

Gergely Orosz surveyed 15 companies, from seed-stage startups to large enterprises, and found the numbers scattered across two orders of magnitude. One seed-stage company went from about $200 per developer per month six months ago to $3,000 today. At a large fintech, individual developers were running $500 a day. A healthcare company watched a single engineer spend $1,400 in one day. At a late-stage infrastructure company, high-end users averaged around $1,000 a week, per The Pragmatic Engineer’s survey.

Responses split almost evenly. About half the companies are running what Orosz calls the “let it rip” approach: no caps, measure the impact, decide later. “Budget is not the concern right now,” one manager told him. The reasoning is defensible arithmetic, since an engineer costs $200,000 to $400,000 a year, and $3,000 a month of tokens is under 15% of that. If the tokens buy back a meaningful slice of that engineer’s time, the trade is obvious.

The other half are pulling levers before the productivity case is proven. The levers, roughly in order of how cheap they are to pull:

• Change the default model. Sonnet instead of Opus for routine work costs nothing to implement and moves the bill immediately.
• Add model routing. Match task complexity to model tier, escalate only on failure.
• Set per-developer caps. Survey respondents mentioned thresholds from $100 to $1,400 a month.
• Turn off high-effort settings for trivial tasks, where extended reasoning burns tokens on problems that never needed it.
• Watch the heavy users, then check whether their output justifies the spend rather than assuming either answer.

The most interesting datapoint in the survey isn’t a dollar figure. One healthcare engineering manager reported that his team is no longer blocked on execution capacity at all. It’s blocked on Product and Design, because engineers now clear tickets faster than anyone can write them. If that generalizes, token spend stops being an engineering line item and becomes an argument about where the actual constraint moved, which is a much harder conversation to have with a CFO than “we need another $50,000.”

Negotiation is thinner than you’d hope. Cursor offers tiered discounts starting around 5% above $1 million in annual spend. Orosz found no reported discounts from Anthropic even at $5 million and above, which tells you which side of this market has pricing power in 2026. That asymmetry is also why a grey market exists. Researcher Matt Lenhard counted 49 relays reselling OpenAI and Anthropic tokens, the cheapest running 97.8% below list, and procurement teams squeezed on price are exactly the audience those relays court. Buying from one means handing your prompts to an unknown intermediary, which is a cost that doesn’t show up on the invoice.

What this means for you

Get a per-developer number before someone in finance asks you for one, because that conversation goes very differently when you already have the figure. Most orgs can’t produce it today, and “we have an Anthropic bill” is not an answer that survives a budget review.

Then pull the cheap levers first. Changing the default model and adding prompt caching costs you a config change and no policy fight, and it typically moves the bill more than any cap you’d have to defend in a team meeting. Save caps for after you’ve measured, because a cap applied to the wrong population, engineers instead of the PDF-to-slides crowd, buys you a slower team and roughly the same invoice.

One forward-looking note worth pricing in: Anthropic says auto mode may increase token consumption, and it becomes the Claude Code default for Pro, Max and Team accounts on August 14. If your spend is already uncomfortable, that lands this week. Check your defaults before Friday rather than reading about it in next month’s invoice.