AI Spend Is a Labor Cost Now :: mamonas.devm.

mamonas.devJul 21, 2026tech7 min read<br>AI Spend Is a Labor Cost Now<br>Last month I ended my Fable 5 post saying I was seeing more and more content about companies starting to look at their AI costs. It didn&rsquo;t slow down after that, so I want to actually go through it this time.<br>A year ago companies were pushing engineers to use AI harder. Uber ran internal leaderboards ranking engineers by Claude Code usage, and consuming more tokens was treated as a good sign.<br>This spring the same companies started installing caps. Uber exhausted its 2026 AI coding budget in four months and now limits engineers to $1,500 a month per tool. Tesla went from leaderboards to a $200 weekly cap. Microsoft reportedly pulled most of its internal Claude Code licenses and pointed everyone at its own Copilot instead.<br>I&rsquo;ve been reading these stories for weeks trying to decide what they mean. The bills are real and the panic makes sense to me, but I think the caps are comparing the bill against the wrong budget.<br>Why the Bills Exploded⌗<br>The easiest way to see where the money goes is to look at one task instead of the whole invoice. When Uber&rsquo;s CTO demoed agentic coding internally, he spent $1,200 in a single two-hour session, with nothing going wrong. That&rsquo;s just what happens when an agent plans, reads files, calls tools, reviews its own output, and retries, with every round carrying the full history of the rounds before it.<br>Part of that cost is fixed before the agent does anything. When someone logged what Claude Code actually sends, every request carried around 33,000 tokens of system prompt and tool definitions before the user&rsquo;s prompt, where lighter harnesses carry around 7,000.<br>Multiply that by an engineering org and the macro numbers start to make sense. Per-token prices have fallen more than 90% since 2023, but total LLM spending doubled since late 2025 anyway. Economists call this Jevons paradox, when something gets cheaper you end up using more of it. An agent using 20 times the tokens at 75% off still comes out to five times the bill.<br>Anthropic&rsquo;s own docs show the same thing. In April it doubled its cost guidance for Claude Code from $6 per developer per active day to $13, and prices hadn&rsquo;t changed, the default model just uses more tokens now.<br>The Seat License Stops Working⌗<br>The pricing changes over the past year all point in the same direction. Cursor repriced Pro around usage credits in June 2025 and had to refund the surprise bills that followed. Anthropic added weekly rate limits for Claude Code power users a month later. This year GitHub Copilot switched to metered AI credits and Databricks Genie went pay-as-you-go, with only a small free monthly allowance per user before billing starts.<br>The reason is the variance between developers. Cursor&rsquo;s own data, via The Pragmatic Engineer, shows the top 1% of developers generating about 45 times the median&rsquo;s output in lines of code. A seat license averages that out across the org, which mostly worked until agents came along. Now one enthusiastic engineer with a harness full of loops can consume what a whole team used to.<br>Ethan Ding called this squeeze in July 2025: demand concentrates on the frontier model, agents multiply tokens per task, and flat subscriptions become a losing bet for whoever sells them. A year later that&rsquo;s pretty much what happened.<br>The $165,000 Rewrite⌗<br>This is the context for the story that made me want to write this post. Earlier this month the Bun team, now part of Anthropic, published a writeup of rewriting the Bun JavaScript runtime from Zig to Rust: 535,496 lines of Zig, converted in 11 days by one engineer supervising about 64 Claude agents at a time. The disclosed usage was 5.9 billion uncached input tokens and 690 million output tokens, &ldquo;around $165,000 at API pricing.&rdquo;<br>That&rsquo;s $165,000 for one pull request, more than ten thousand days of Anthropic&rsquo;s own $13-per-developer guidance. If that number landed on a CFO&rsquo;s desk as a software line item, it would trigger exactly the kind of cap Uber and Tesla just installed.<br>Jarred Sumner, Bun&rsquo;s creator, estimates that by hand the rewrite would have taken &ldquo;3 engineers with full context on the codebase about a year.&rdquo; A Hacker News commenter did the salary math, at Bay Area compensation that&rsquo;s &ldquo;at least $1.5M in compensation alone,&rdquo; and it would have taken a year instead of eleven days. Against that, the same $165,000 is roughly a 90% discount, done about 30 times faster.<br>And it wasn&rsquo;t a demo, the port passed 100% of Bun&rsquo;s test suite on all platforms before merging, and Claude Code itself has been running on the Rust version since mid-June, which Simon Willison verified by finding the Rust source filenames embedded in the shipped binary. Whether that bill is outrageous or a bargain depends entirely on which budget you put it against.<br>The...