Uber exhausted its 2026 AI budget by April after engineers rapidly adopted agentic coding tools. Bloomberg reported the company has imposed a $1,500 monthly cap per employee and per agentic coding tool, including Anthropic's Claude Code and Cursor, with overspend allowed only with explicit permission and usage shown on an internal dashboard. Forbes reporting showed the rollout delivered measurable output: by spring most engineers were using AI tools monthly and a large share of committed code originated from those tools, but token-based billing sent invoices far beyond forecasts. Anthropic is changing its pricing model on June 15, a shift that will make heavy agent use more expensive for companies that do not curb consumption.

Adoption leapt from about 32% of engineers in February to 84% classified as agentic coding users by March, yet token metering turned those gains into runaway invoices.

How a productivity win became a pricing problem

Uber began rolling Anthropic’s Claude Code broadly to its engineering organisation in December 2025. Internal metrics reported to Forbes showed usage climbing steeply over the first months of 2026. By spring, roughly 95% of engineers were using AI tools on a monthly basis, and about 70% of the code that teams committed traced back to suggestions or generation from those tools. The same reporting noted that around 11% of live backend updates at that point were produced by autonomous agents operating with no human in the loop.

From a productivity viewpoint, the rollout was successful. But token-based billing, where cost scales with the compute and text tokens consumed, turned productive runs into surprisingly large monthly bills. Forbes reported that typical monthly spend per engineer ranged from about $150 to $250, while heavier users averaged $500 to $2,000. Praveen Neppalli Naga, Uber’s chief technology officer, told colleagues that he spent $1,200 during a two-hour demo, according to Forbes. The combination of heavy use and granular, variable pricing exhausted Uber’s planned AI allocation for 2026 in roughly four months, Naga disclosed in April.

The blowout stood out inside Uber’s accounts because the company’s research and development outlay is already large: total R&D spending reached $3.4 billion in 2025. That scale made the budget miss look less like a growth issue and more like a pricing and cost-model failure, finance teams told internal meetings cited in reporting.

A cultural incentive structure accelerated the problem. Forbes reported that engineering teams were ranked on internal leaderboards by Claude Code usage, a competitive pressure that encouraged heavier token consumption. That metric, useful for measuring adoption, was not aligned with the behaviour finance had modelled when it set the annual AI envelope.

Mastheads

The result was an organisational gap: engineering incentives pushed for consumption, while budget controls and forecasting assumed steady per-seat costs rather than highly variable token consumption from agent orchestration.

Token metering differs from traditional per-seat licence models because the same engineer and the same tool can consume dramatically different quantities of tokens. Simple autocompletes incur modest charges. Orchestrating parallel agents across a monorepo, or running continuous agentic workflows, multiplies token use and, with it, invoices.

Bloomberg reported the immediate management response in early June: Uber has set an internal limit of $1,500 per month per employee and per agentic coding tool, covering Claude Code and Cursor among others. The rule allows teams to exceed that ceiling, but only with explicit permission and when the excess is visible and justified on a central dashboard. The new cap is intended to add a financial gate to a process that had until now been governed mainly by technical adoption and operational enthusiasm.

The cost issue has been sharpened by changes at vendors. Anthropic announced on May 13 that it would introduce a separate monthly credit meter for agent tools and third-party harnesses, and that those usages would be billed at full API rates from June 15. That structural pricing change raises the cost of heavy agent orchestration and will increase the pressure on companies that have already scaled agentic use without corresponding budget controls.

Outside Uber, reporting has shown that some firms are reassessing tool choices. Several reports suggest Microsoft has cancelled many direct Claude Code licences and is directing some engineering teams toward GitHub Copilot CLI, a move that reflects the wider market reaction to rising agent costs and shifting vendor terms.

Internally, the company’s senior managers have acknowledged the trade-offs. Andrew Macdonald, Uber’s president and chief operating officer, said on the Rapid Response podcast that "it’s very hard to draw a line" between rising Claude Code use and the delivery of consumer-facing features. On an earlier earnings call, CEO Dara Khosrowshahi said that about 10% of the company’s committed code is built by autonomous agents. Those comments underline that the productivity gains are real and tied to product outcomes, which complicates any effort to throttle usage without affecting delivery.

Uber’s corrective steps try to strike a balance. The $1,500 cap and the requirement for visible approval create friction where none existed, nudging teams to think about token efficiency and to route heavy workloads through architecture that reduces cost. But the policy also leaves room for exceptions, which reflects the competing priorities: short-term delivery of features and the mid-term need to bring cloud and API costs under control.

Related Articles

Anthropic's separate agent metering takes effect on June 15, the first clear test of whether Uber's $1,500 cap and approval process will prevent a repeat of April's budget exhaustion.

This article was created with AI assistance.