How Uber Kept AI Spending Flat as AI-Agent Usage Rose 9.4×

Uber’s AI spending initially ran beyond its 2026 budget, but the company did not respond by simply throttling access. Instead, it treated AI as an engineering-cost optimization problem: match each workload to an appropriate model, limit excessive context, reuse cached prompts, and show engineers what their sessions cost. 1

That operating model helped keep total AI spending broadly stable from April while weekly AI-agent requests increased 9.4 times from February. Using the same AI model, Uber reported that cost per 1,000 requests fell nearly 34% from its April peak, while cost per session declined 52% from its June high. 1

Uber’s AI usage kept growing

The cost improvements came as adoption expanded rather than contracted. AI agents accounted for more than 70% of code-change submissions, engineers ran over 30,000 agent tasks daily, and the number of employees using AI tools more than quadrupled. 1

That combination—much higher usage with broadly stable spending—points to a fall in unit cost. It does not mean the workloads became free; it means Uber changed how tokens and model capacity were consumed.

1. Routing tasks to fit-for-purpose models

Uber sends work to the model that can handle the task at the best cost, rather than using the most capable or expensive model by default. Simpler tasks can be directed to lower-cost models, while more demanding work receives greater model capacity. The company is also evaluating alternatives, including open-weight models, for specific use cases. 1

This makes model selection a workload decision instead of a one-size-fits-all setting. At scale, even modest savings on routine requests can materially change an organization’s average cost per task.

2. Limiting long-running agent context

Uber caps interactive sessions at 400,000 tokens, even when a model supports a context window of up to one million tokens. 1

The limit addresses one of the cost risks of coding agents: a session can accumulate files, tool results, instructions, and repeated interaction history. Without a boundary, exploratory or poorly controlled sessions can keep processing increasingly large amounts of context.

A token cap does not eliminate agent use. It creates a predictable ceiling for the most context-intensive sessions and gives engineering teams a control point for managing usage.

3. Extending prompt-cache life

Uber extended prompt-cache duration from five minutes to one hour. The change was intended to better match engineers’ working patterns, in which an AI session may sit idle for longer than five minutes before work resumes. 1

Longer cache reuse can prevent the same context from being processed repeatedly. For agentic coding workflows, where the same repository or instructions may be revisited across a session, that can reduce unnecessary token processing without changing the task itself.

4. Showing engineers the running cost

Uber gives engineers real-time visibility into the cost of an AI session in their terminals. 1

This turns usage into an immediate engineering feedback loop. Instead of discovering an expensive workflow only through a later finance report, an engineer can see the cost while experimenting and adjust the session, model, or context when necessary.

The measure also places cost awareness inside the product workflow, where decisions about prompts, retries, context, and model choice actually occur.

What Uber’s results suggest for enterprise AI

Uber’s experience illustrates a broader operating principle: controlling enterprise AI spending is not only a budgeting exercise. It is also a systems-design problem.

The most durable controls in this example are:

• Workload-based model routing: use less expensive models where their capabilities are sufficient.
• Context governance: set limits for sessions that can accumulate large amounts of token-heavy history.
• Prompt caching: reuse repeated context instead of processing it from scratch.
• Unit-cost telemetry: make the marginal cost of an AI workflow visible to the people running it.
• Model competition: evaluate alternatives, including open-weight models, rather than relying on one default for every task. 1

These controls are particularly relevant to AI agents because agentic work is iterative. A single user request can lead to multiple rounds of planning, tool use, verification, and revision, making total token consumption less predictable than a one-shot interaction.

Uber’s reported results should therefore be read as an example of operational efficiency, not proof that rapidly growing AI usage carries no financial risk. The company kept spending broadly stable only after changing the economics of each request and session. For other enterprises, the lesson is straightforward: scale adoption and cost governance together, with measurement built into the tools from the start.