Uber holds AI costs steady with cheaper models, employee spending oversight

Key takeaway

• Uber kept its artificial intelligence spending stable in the second quarter by switching to cheaper AI models for some tasks, improving default settings, and letting employees track their own AI spending more closely, CFO Balaji Krishnamurthy said Wednesday.

• The company had previously burned through its 2026 Claude Code budget in just a few months, making cost control a priority.

• Uber is now using AI to make incremental improvements across its app and operations—such as predicting rider destinations and helping customers build grocery shopping carts—rather than placing large bets on single features.

3 Key Points

FAQ

How did Uber cut its AI token spending per unit?

Uber set better defaults for different use cases, used lower-cost models for some tasks, and gave employees clearer visibility into their AI spending, CFO Balaji Krishnamurthy said. This combination allowed cost per token to decline over several months even as adoption increased.

What is an example of Uber's AI features in use today?

One example is Uber's destination-prediction feature: when customers open the app to request a ride, the app suggests where they are going; Khosrowshahi said the app correctly predicts the destination three-quarters of the time. Another is Cart Assistant in Uber Eats, which creates a shopping cart from a recipe or shopping list and has driven cart sizes that are twice as large as those of users who did not use the tool.

Why did Uber run out of its Claude Code budget so quickly?

The article does not specify the cause of the rapid budget depletion. CTO Praveen Neppalli Naga went viral earlier this year after saying Uber had spent its Claude Code budget for 2026 after just a few months, but no explanation is provided in the article.