Coding Agents Hit the Corporate Expense Account

Uber caps coding agents at $1,500/month per tool per engineer while Anthropic claims Claude now ships 8x more code than human baselines. The price of an AI pair programmer is becoming a line item, not a rounding error.

This week the question shifted from can AI write production code to how much should we budget per engineer to let it. Uber reportedly capped coding-agent spend at $1,500/month per employee per tool, and Simon Willison notes Brex is at $2,000/month — meaning a developer running Cursor, Copilot, Claude Code, and Codex in parallel could legitimately burn $6,000 in tokens a month. That is not a side-project number. That is a senior engineer's salary cut, paid to inference providers.

The spending is being justified by escalating capability claims. Anthropic published internal data on June 4, 2026 arguing that Claude is now accelerating its own development, with Mythos Preview improving on human research decisions 64% of the time (up from 22% in 2024) and hitting a 52x speedup on a standard model-training optimization task — a job that takes a skilled human 4-8 hours. Anthropic also reports its own engineers ship 8x more code per quarter than they did in the 2021-2025 baseline. These are vendor-reported numbers on vendor-designed benchmarks, and they should be read that way. But the direction of travel is consistent with what practitioners are seeing in the field.

The tooling layer is racing to capture that spend. OpenAI expanded Codex with plugins covering 62 apps and 110 skills across sales, design, and data analytics — a clear push to turn Codex from a coding assistant into a horizontal work agent. GitHub shipped a dedicated Copilot desktop app for managing parallel agent workstreams, and Windsurf relaunched its IDE as Devin Desktop with an Agent Command Center for multi-agent orchestration. Everyone is converging on the same product shape: a control plane for fleets of coding agents.

For builders, the implication is uncomfortable. As El Bruno documented this week, the move from flat-rate Copilot subscriptions to usage-based billing means the muscle memory of "just ask the agent" now has a meter attached. For buyers, the Uber and Brex caps are an early read on what enterprises think a coding agent is actually worth — somewhere between an IDE license and a junior engineer's tools budget. For creators of these tools, the window to land inside that cap before procurement sets it in stone is closing fast.

The bottom line: Set your team's per-engineer agent budget this quarter, before finance sets it for you based on last month's surprise invoice.

Sources