Uber Caps AI Coding Spend at $1,500 After Its 2026 Budget Ran Out

Uber has put a $1,500 monthly limit on each employee’s use of an agentic coding tool after its 2026 AI coding budget was exhausted by April. The immediate cost control addresses a visible spending problem, but Uber’s president and COO said the company had not yet shown that heavier use was producing better products for riders and drivers.

The episode captures a growing enterprise problem: token consumption is easy to count, while the value of the work is harder to isolate. Gartner expects spending on AI agent software to approach $207 billion this year, up from $86.4 billion in 2025. But the same developer can generate very different costs depending on whether they use simple suggestions or run several agents on a complex migration.

The first responses do not agree

Some companies have pulled back. Microsoft questioned the cost of Claude Code licenses before canceling them across its Experiences and Devices division. Duolingo reversed a plan to factor AI use into performance reviews after employees objected to being pushed to use the tools for their own sake.

Others argue that the wrong models, rather than too much use, drive the bill. Agiloft removed its caps, made lower-cost models the default and added escalation to frontier models only when needed. It also uses infrastructure-level routing and caching, shifting the choice away from every individual prompt.

SUSE has taken a people-led route for now, sorting work into Daily Work, Autonomous Agents and one-time “Curve Jumping” efforts. Model decisions still sit with individual developers, though the company plans a proxy layer to help manage them later.

Routing moves the choice into the platform

Databricks’ beta Smart Routing in Unity AI Gateway offers two paths: recommend a model or automatically route a request. Databricks says the system considers signals including prompt intent and length, attached files, stack traces, change scope, reasoning depth and execution complexity. For agentic jobs that can branch into many model calls, it evaluates at execution boundaries rather than every call.

Unity Gateway budgets can set shared or per-user thresholds, alert administrators, or block further requests. They can also include external-model spending in beta. Those are controls, not a precise billing guarantee: enforcement relies on near-real-time estimates, so final spend can land above or below the configured threshold.

Savings need a business-side test

A company can get almost every engineer using AI and still be unable to prove the spend produced more features shipped, more bugs fixed or more customer problems solved. Everlaw is an exception in the examples cited: it reported spending $3,500 in tokens on a Java infrastructure effort that reduced implementation from 9.5 engineer-months to 2.5.

The same company also found a limit to automated output: it discarded agents’ work porting interface code from Dojo to React because the frameworks used different assumptions about state and view. Its revised method asks an agent to document the old system’s behavior before attempting a rebuild.

The next move is therefore not simply to reduce token use. Teams still need a way to detect whether lower-cost routing changes quality in their own domain, particularly as models change. Promova is targeting a 40% Opus, 50% Sonnet and 10% Haiku mix, but treats that mix as a direction rather than a durable endpoint; Opus with a one-million-token context window had represented about one-third of its monthly spending.