Microsoft confirmed it with its own research: engineers using Claude Code merged 24% more PRs.
Then on June 30, 2026 — the last day of the fiscal year — they cut licenses for more than 5,000 engineers.
The results were proven. So why'd they turn it off?
But the bill looked off
Microsoft's Claude Code rollout started smoothly. They rolled it out to tens of thousands of engineers in December 2025, and the gains held steady for four straight months. Engineers who adopted it merged 24% more PRs per day on average than those who didn't.
The catch: it started as a flat per-seat plan. Nobody could see usage, so everyone used it freely — and once billing switched to usage-based, the real cost surfaced. Spend climbed to $500–$2,000 per employee per month, and by June 30, the end of the fiscal year, they'd cut licenses for a big chunk of engineers.
Around the same time, Uber blew through its entire 2026 AI coding budget in just four months. Their fix was simple — a $1,500-per-employee-per-model monthly cap.
Hacker News developers reacted pretty coldly to the news. Nobody knows what a request is going to cost until it's already run — basically signing a blank check every time.
Everyone's hitting the brakes
This isn't just an Uber-and-Microsoft thing. Caps are showing up across every industry.
| Company | Policy |
|---|---|
| AhnLab | 3,000 GitHub Copilot credits/month cap |
| LG Electronics | Base token budget issued, overages need pre-approval |
| Samsung Electronics | "Token ROI" system measuring output per token spent |
| Netmarble & NHN | Monthly usage caps tiered by role and team |
| Hyundai Motor | No duplicate subscriptions, only approved AI tools allowed |
| Uber | $1,500/employee/model monthly cap |
| Walmart & Amazon | Cheaper models set as default |
You can see why the mood shifted. A Gartner senior analyst warned that AI token spend could match or exceed a software engineer's monthly salary within two years. Some extreme cases are already running $20,000–$32,000 a month.
At Meta, one employee burned through 280 billion tokens in a single month. Company-wide, that adds up to 60 trillion tokens in 30 days. At this point, "productivity" and "spend" become two completely different questions.
But switching to the cheapest model isn't the fix either
There's a common misconception here: "just use the cheapest model, problem solved." But that's not actually how companies that manage this well are doing it.
Databricks wrote about this and argued the shift should be from the "intelligence frontier" to the "efficiency frontier" — instead of defaulting to the newest, biggest model, pick whichever model gives you the best price-to-performance for each specific task.
The blog even cites Stripe, which tested upgrading to a top-tier model (Opus 4.7) and ultimately passed — the extra cost didn't deliver enough extra value to justify it.
| Using it unmanaged | Managing with routing + caps | |
|---|---|---|
| When you find out the cost | After the bill arrives | Real-time dashboard |
| Model choice | Always defaults to the priciest | Auto-routed by task difficulty |
| When you exceed budget | Unlimited or a full shutoff | Alert at 75%/90%, then downshift |
| Real-world result | Uber burned a year's budget in 4 months | ~50% token reduction via harness + caching tweaks |
Anthropic is following the same playbook. On July 2, 2026, they added spend threshold alerts (75%/90%) and per-team model defaults to Claude Enterprise. There's also an analytics chat feature now — admins can just ask which team doubled its usage this month and get an instant answer.
5 things you can fix right now
- Make usage visible first
Start with the admin dashboard you probably already have. Claude Enterprise Analytics, Cursor Business reports — these features ship built-in. - Split models by task difficulty
Default to a cheap model for simple tasks, and only escalate to a high-power model for complex refactors. Using the cheapest model for everything just means more rework — and more tokens in the end. - Set alert thresholds first
Instead of a hard cutoff, design it so you get alerted at 75%/90%, then automatically downshift to a cheaper model from there. - Cut token waste
Compress context regularly, break tasks into smaller chunks, and turn on prompt caching. Databricks cut its token usage nearly in half with just these tweaks. - Keep your codebase lean
Hacker News developers point out that the bigger a codebase gets, the more agents wander and burn tokens. A minimal codebase is basically the prerequisite for running an AI agent well for more than a month.
If your team is small
Unless you're operating at Uber or Microsoft scale — thousands of people — a dedicated AI gateway is probably overkill. Start by turning on the admin console (usage dashboard, alerts) in the tools you're already using.
Want to dig deeper
The original AI coding cost management piece Databricks' four strategies covering the efficiency frontier, routing, and budget management databricks.com
The Microsoft rollout study The original paper tracking tens of thousands of engineers over 4 months, validating the 24% PR increase arxiv.org
Claude Enterprise's spend management features Anthropic's own writeup on setting alert thresholds and model defaults claude.com
Real-world corporate token budget examples A newsletter roundup of actual cap numbers from Meta, Uber, Stripe, and more newsletter.semianalysis.com
How Korean companies are capping AI tokens A rundown of how AhnLab, LG Electronics, Samsung Electronics, and others are responding insightkorea.co.kr
The original Hacker News thread Raw developer reactions and pushback news.ycombinator.com
