Three months ago, Uber was a cautionary tale. Fortune reported in May that the company burned through its entire 2026 AI budget in just four months, prompting its COO to publicly question whether it was worth it.

The story is very different now.

According to an Axios exclusive published August 27, Uber’s weekly AI agent requests have grown 9.4 times since February 2026. Weekly active users across all agentic tools grew 7 times over the same period. And yet Uber’s total AI spending has stayed stable since April.

That’s not a typo. Dramatically more usage. Roughly the same cost.

The company’s CTO, Praveen Neppalli, shared the details with Axios. The numbers are striking: the cost per 1,000 requests is down about 34% from its April peak, and the cost per single session has fallen 52% from its June high. AI agents now handle more than 70% of code-change submissions at Uber, and engineers run over 30,000 AI-agent tasks every day.

The number of people using AI tools inside Uber has more than quadrupled.

How They Actually Got There

The turnaround wasn’t magic. Uber made a set of practical, replicable changes to how it manages AI workloads.

Model routing: Not every task needs the most expensive model. Uber now routes simple tasks to cheaper models and reserves frontier AI for genuinely complex problems. The company is also actively testing open-weight models as cheaper alternatives for specific use cases.

Prompt caching: Uber extended its prompt cache window from five minutes to one hour. The reason is straightforward: engineers often leave sessions idle for longer than five minutes, and every idle restart was resetting the cache and triggering unnecessary costs.

Session caps: Uber limits interactive sessions to 400,000 tokens even when using a model capable of handling one million. This alone significantly reduced runaway spend from sessions that never actually needed more context.

Real-time cost visibility: Engineers now see what their AI usage is costing them in real time. When people can see the meter running, they use resources more thoughtfully.

These are not exotic interventions. They’re operational discipline applied to a new category of spend.

What This Means for Business

The May story about Uber blowing its AI budget became a headline because it confirmed what a lot of executives suspected: AI agents are expensive, and the costs are hard to predict.

The August story is more important, and it’s getting less attention.

What Uber demonstrated is that the cost cliff is real but crossable. The path is not to pull back on AI adoption. It’s to get serious about how you’re using it. Model selection, session management, cache tuning, and visibility are the levers. None of them require a PhD in AI.

The broader implication for any business deploying AI agents: the early months will likely overshoot your budget. That’s normal. The companies that come out ahead are the ones that treat AI spend like any other engineering cost, build instrumentation into the workflow, and continuously optimize.

Uber’s 30,000 daily AI-agent tasks and 70% agent-handled code changes are not the story of a company that throttled AI to save money. They’re the story of a company that learned to run AI efficiently at scale.

The difference between “we burned through our AI budget” and “we run 9.4 times more AI work for the same cost” is operational maturity, not a different technology.

For businesses building AI agent workflows today, this is the playbook worth studying: deploy broadly, instrument everything, route intelligently, and optimize continuously. The cost curve bends if you put the work in.

---

Enterprise DNA helps businesses build and deploy AI agent operations across their organizations. If you’re working through the ROI and cost questions around AI adoption, book a discovery call to talk through what that looks like in practice.