Uber used up its entire 2026 AI budget by April, four months into the calendar year, after Anthropic’s Code Claude spread to about 5,000 engineers faster than the company’s financial models expected. Chief Technology Officer Praveen Neppalli Naga confirmed the exceedance to The Information, saying the company is back on plan for its affairs. Uber’s total R&D spend has arrived $3.4 billion in 2025a 9 percent year-over-year increase, which makes the budget collapse less about scale and more about a pricing model that enterprise finance teams haven’t learned how to manage.
The revelation landed alongside a structural change from Anthropic itself. On May 13 the company was announced that paid Claude subscribers will soon face a separate monthly credit meter for dealer tools and third-party braids, which will be billed at full API prices starting June 15. Read together, the two events describe a single problem. Token-based consumption pricing doesn’t behave like the software line items CFOs know how to model, and the gap between what engineers consume and what finance teams expect is no longer hypothetical.
How a coding tool went over a budget
Uber rolled out Claude Code to its engineering organization December 2025. Adoption increased from 32 percent of engineers in February to 84 percent classified as coding agent users by March. until spring, 95 percent of Uber engineers used AI tools monthly, and about 70 percent of the committed code came from those tools. About 11 percent of live backend updates were written by agents without a human in the loop, according to Uber’s own disclosures.
The numbers behind the spending are what make the story instructive, not anecdotal. Monthly costs per engineer ranged from $150 to $250 on average, with power users running between $500 and $2,000. Naga himself mentioned spending $1,200 in a two-hour session during a personal demonstration. The tool did not fail and the engineers did not abuse it. They used it for exactly the workloads it was designed to handle, running parallel agents, rebuilding large-scale codebases, automated test generation, and backend code generation. From a productivity perspective, growth has been successful. From a financial point of view he was a fugitive.
Uber reinforced the momentum by ranking engineers on internal leaderboards based on their use of the Claude Code. This created a cultural incentive to consume more brands, which translated directly into faster budget burn. The teams driving the adoption were not the same teams managing the spending, and this organizational gap turned out to be the fatal flaw.
Why Token Billing Disrupts Traditional Budgeting
Claude Code is not priced per seat. It measures tokens consumed in model calls, meaning that an engineer running autocomplete suggestions consumes a fraction of what an engineer orchestrating parallel agents in a monorepo would consume. The same tool, the same engineer, the same work day, can produce wildly different invoices depending on the workflow chosen. Annual budget cycles based around predictable costs per license cannot absorb this variation.
Microsoft has taken the opposite approach Microsoft 365 Copilot Enterprisewhich sells for $30 per user per month with an annual commitment. The price caps vendor caps and gives funding teams a flat line item they can multiply by headcount. Anthropic’s consumption model gives the seller unlimited upside to heavy users and gives almost no visibility to funding groups. Both models are defensible, and neither is right for every workload, but treating them as interchangeable in a programming cycle is what drove the Uber result.
GitHub is moving Copilot to a credit-based system on June 1 and analysts cited by InfoWorld expect Most vendors will introduce separate consumption groups for dealers and tool use in the next 12 to 24 months. The vocabulary will vary, credits, requests, messages or computing units, but the direction is set. The foregone conclusion of unlimited agent workloads was never going to survive the math, and Anthropic’s announcement in May is the first major confirmation that sellers will pass on cost drivers to buyers rather than absorb them.
The limits of the productivity defense
The typical industry response to consumption-cost stories is that AI pays for itself in productivity gains. The case of Uber complicates this argument. The marginal productivity gain from a senior engineer running agent workflows must overcome a much higher contract cost hurdle than the gain from an engineer running autofill. Five to twenty times increase in consumption per developer documented in agent modeand no public benchmark displays a corresponding multiplier on the output value. Productivity savings also don’t appear on the same line item as AI costs, meaning finance teams can’t offset them in a quarterly review.
There are also operational limits that make the simple cost versus production framework incomplete. Only 43 percent of organizations have formal AI governance policies, according to research data reported in covering Uber’s excess and only 21% have mature governance. Most enterprises are not yet applying to AI the cost control tools that DevOps teams routinely apply to cloud computing. This includes per-engine caps, real-time monitoring of token consumption, and budget alerts before they are exceeded rather than after. Uber deployed Claude Code across the organization without these controls, and the result was visible within a quarter.
What CFOs should take from this
Uber’s experience produces a short list of practical implications for financial leaders seeing their own engineering organizations adopt agent coding tools. The first is that pilot economics do not predict economies of scale for consumption-priced tools, because pilots run on a few engineers using autofill while production runs on entire teams that outsource multi-step workflows to agents. The second is that incentive structures matter as much as pricing. Leaderboards and adoption goals drive token consumption, and any disposition that rewards unfettered use should be framed as an unfettered obligation until proven otherwise.
The third is structural. Anthropic’s June 15 credit change signals that subsidized programmatic usage in subscription plans is ending across the industry. Businesses that built their forecasts on Claude Code’s financials will see their actual cost per unit rise, and the same logic will apply to other suppliers as they follow Anthropic’s lead. Procurement groups that want predictability will have to negotiate fixed-price cost commitment agreements instead of consumption pricing, and the leverage they have in those conversations will depend on whether their engineering organizations have usage caps at all.
Uber isn’t slowing down on its AI push. Naga plans to pilot OpenAI’s Codex with Claude Code, and the long-term vision he describes is one where engineering agents handle coding, testing, and development with humans acting as orchestrators. This direction is consistent across large engineering organizations that are now adopting these tools. The open question for boards is not whether to deploy them, but whether finance functions have visibility into their costs when engineers stop holding back.
