Uber's AI Coding Costs Soared Because Engineers Always Used the Priciest Models

Uber exhausted its 2026 AI coding budget within four months, with one executive spending $1,200 in a single two-hour Claude Code session and heavy users reaching $2,000 per month. The company responded by imposing a $1,500 per-engineer monthly spending cap, but critics argue this treats the symptom rather than the root cause. Analysis of AI coding usage suggests only about 15–20% of tasks genuinely require expensive frontier models, while the remaining 80% can be handled adequately by mid-tier or lightweight models. A model-routing approach — directing complex architectural work to frontier models and routine tasks like boilerplate or unit tests to cheaper alternatives — reportedly cut one team's monthly API bill from roughly $10,000 to $3,000 without sacrificing output quality. Spending caps risk backfiring by pushing engineers to conserve budget on simple, high-value tasks while overspending on complex ones, making intelligent task-level routing a more effective long-term solution.