Uber exhausts AI coding budget in four months

Uber CTO Praveen Neppalli Naga disclosed that the company exhausted its entire 2026 AI coding tools budget in just four months, following the adoption of Anthropic's Claude Code by approximately 5,000 engineers. Adoption rates climbed from 32% in February to 84% by March, reaching a 95% monthly active rate by spring. The average cost per engineer was $150-$250 monthly, with some power users reaching $500-$2,000; Naga himself spent $1,200 in a single two-hour session. Salesforce CEO Marc Benioff announced that Salesforce will likely use $300 million of Anthropic this year across an engineering force of about 15,000. Microsoft's Experiences + Devices division, which includes Windows, Microsoft 365, Outlook, Teams, and Surface, is reportedly cancelling direct Claude Code licenses by June 30, redirecting engineers to GitHub Copilot CLI, as reported by The Verge's Tom Warren and confirmed by Windows Central. LangChain has introduced a four-stage operational framework to manage coding agent spend before token burn becomes structural overhead.

The issue stems from "tokenmaxxing"—using token burn as a productivity proxy—and tool fragmentation, which obscures burn rates. Features may pass through Claude Code, Cursor, and GitHub Copilot Chat, each emitting incompatible telemetry. LangChain's blog post notes that native dashboards fail with multiple tools; cost comparison without normalization is impossible, leaving teams unable to track spending.

LangSmith, the first stage, ingests OpenTelemetry spans from GitHub Copilot, session hooks from Cursor and OpenCode, and extension data from Pi, normalizing them into a single trace model. Sessions are queryable by thread_id, model, provider, or tool name, exposing user-assistant turns, token usage, cost, shell commands, MCP activity, and subagent invocations. This normalization is essential, as tool calls in different platforms are recorded differently, and LangSmith standardizes them for clean data processing.

The pipeline then moves to standardization, optimization, and governance. LangSmith attributes cost per session across tools, allowing teams to compare Cursor and Claude Code on identical workflows in dollars. LangChain Engine analyzes traces to identify redundant tool calls and recommends consolidation. The LLM Gateway enforces cost caps at user, team, and org tiers and routes workloads to open-source models when appropriate. LangChain posits that open-source models are now capable enough for scoped subtasks, reserving frontier models for necessary tasks.

At Uber, the overrun was compounded by internal leaderboards that ranked engineers by Claude Code usage, creating a direct incentive to maximize token burn, according to Forbes. The company has since instituted a $1,500 monthly cap per employee per agentic coding tool, as reported by TechCrunch. Salesforce told AI Magazine that AI now accounts for 30–50% of its overall workload and that its engineering org is roughly 30% more productive after integration of broader AI tooling including Agentforce and other tools, not solely Anthropic token spend, while Uber says roughly 70% of its committed code originated from AI tools, with 11% written entirely by agents and the remainder reviewed by humans, per Forbes. Those gains are only defensible if tied to shipped features rather than aggregate token volume. Uber COO Andrew Macdonald told Fortune that the link between rising AI spend and useful consumer features "is not there yet."

The integration pattern leverages existing tooling by tapping into Copilot's OpenTelemetry emissions and Cursor's hooks, then adding governance. However, normalization requires every tool to emit the required telemetry; if a subagent or MCP server hides token counts, the cost trace fails. The open-source fallback assumes the existence of eval harnesses to prove the cheaper model does not degrade output quality, which most teams lack. Gateway-level governance introduces latency, impacting developer productivity.

Treat coding agents as distributed inference workloads requiring unified observability before optimization, because measuring across incompatible vendor formats is prerequisite to keeping frontier-model burn from outpacing engineering value.