Uber's AI Agent Requests Grew 9.4x Since February While Total Costs Stayed Flat

My take: In four months, Uber burned through its entire annual AI coding tool budget and had to reset from scratch. Six months later, the same company grew its agent request volume 9.4 times since February, while keeping total spending flat since April. Cost per thousand requests is down 34% from its peak, and cost per session is 52% lower than its June high.

What changed was not the model: it was the strategy. Uber redirected routine tasks to less expensive models, improved prompt caching, and began exploring open-weight models for simpler work. The result: more than 30,000 agent jobs per day and over 70% of code changes generated by AI, with the bill no longer climbing.

Uber's CTO said it plainly: "The next phase will not be about who spends the most tokens, but who uses them most efficiently."

Is your team still in the tokenmaxxing phase, or are you already optimizing your AI spend per result you actually get?