The Unlimited AI Era Is Over. GitHub Copilot Just Made It Official.

One company spent $500 million on Claude in a single month. Microsoft canceled its own engineers' licenses. Uber burned its entire 2026 AI budget by April. Here is what is actually happening and what to do about it.

• On June 1, 2026, GitHub Copilot changed how it charges you. Instead of a flat monthly subscription, every Copilot plan now runs on a token credit system. Every prompt, every code completion, every file review, every agentic session draws from a monthly credit pool based on how much compute it actually consumed.
• The base prices did not change. Copilot Pro is still $10/month. Pro+ is still $39/month. Business is still $19/user/month. Enterprise is still $39/user/month. What changed is what those prices actually buy you. The flat-rate unlimited era is over.
• The backlash on GitHub's community forum was immediate. One developer reported that two prompts on June 1 burned through their entire month's credits. Another reported their session on May 31 felt predictable and normal. The identical session on June 1 consumed dramatically more of their quota.
• This is not a GitHub problem. It is an industry problem. And it has been building for two years.

The $500 Million Claude Bill That Explains Everything

In late May 2026, Axios published a story that stopped a lot of finance departments mid-meeting. An AI consultant revealed that one of their unnamed enterprise clients had accidentally spent $500 million on Anthropic's Claude in a single month.

No malicious intent. No runaway system. Just a company that gave all its employees unrestricted API access to Claude without spending caps, usage limits, or any dashboard tracking who was consuming what.

Employees used Claude for everything. Engineers ran complex agentic coding workflows. Analysts processed large documents repeatedly. Some reportedly used the tool for tasks as simple as checking the weather. Because access was uncapped, the billing was too.

The unnamed company is not alone. Microsoft canceled its own internal Claude Code licenses for thousands of engineers that same month, after realizing costs were running between $500 and $2,000 per engineer per month. Uber burned through its entire 2026 AI budget by April. Amazon shut down an internal AI usage leaderboard after discovering employees were running pointless low-value prompts to inflate their usage numbers.

These are not small companies making amateur mistakes. They are among the most sophisticated technology organizations on earth. That is the point.

Why Flat Rate AI Was Always a Subsidized Fiction

When GitHub Copilot launched at $10/month in 2022, the AI model underneath it was cheap to run. GPT-3.5 class inference costs were low. Most users generated short completions. The economics worked.

Over the past two years, everything changed. Users moved from short completions to multi-file agentic sessions that can run for hours. Context windows expanded from 8,000 tokens to one million tokens. The underlying models got dramatically more capable and dramatically more expensive to run.

A heavy Copilot user in 2026 consumes 50 to 100 times more compute per session than a typical user did in 2022. That usage was being subsidized. The flat-rate model was effectively transferring cost from heavy users to light users and to the platform's own P&L.

GitHub's announcement framed usage-based billing as alignment between pricing and actual cost. That framing is accurate. The uncomfortable truth is that flat-rate unlimited AI was always a growth-phase subsidy, and the growth phase is over.

The Team Credit Pool Problem

For business and enterprise teams, there is a specific operational problem that has not received enough attention.

Credits are pooled at the organization level. If one engineer on a team of five runs a heavy agentic coding session that processes large files repeatedly across a three-hour sprint, they can exhaust a significant portion of the team's monthly credit allocation before lunch.

Other team members then hit usage limits mid-session, mid-pull-request, mid-review.

This is new behavior. Under the old premium request model, individual usage was tracked per seat with clearer per-user limits. The token credit pool model creates shared exposure to heavy users in a way that is harder to predict and harder to control.

GitHub has stated that admins can set spending limits and usage caps at the organization level. Those controls exist. But most teams do not have someone actively monitoring AI credit consumption the same way they would monitor cloud compute costs. That gap is where the surprises live.

The Alternatives Have the Same Underlying Problem

After June 1, Cursor and Windsurf saw a noticeable uptick in sign-ups from developers migrating away from Copilot. Both are excellent tools. Cursor's Agent mode in particular handles complex multi-file tasks with a quality that continues to impress.

But here is the structural reality: every AI coding tool faces the same economics.

Cursor's $20/month Pro plan uses Claude Sonnet and GPT-4o under the hood. Windsurf's $15/month Pro plan runs on similar frontier models. Kilo Code at $55/month uses Claude models. Every one of these plans is priced based on the platform's calculation of average usage. When your usage is above average, the economics stop working for them.

The difference today is that Cursor still offers a flat-rate plan. That could change. GitHub was flat-rate until it was not. The structural pressure pushing platforms toward usage-based pricing does not disappear because a competitor chose a different model today.

What Companies Are Learning the Hard Way

The companies that ran into six and seven figure AI bills in 2026 share a common pattern. They treated AI tools like traditional SaaS subscriptions. Flat fee. Predictable. Seat-based.

AI billing does not work that way. It works like cloud compute. The cost scales with what you actually do, not with how many seats you paid for.

Amazon learned this when employees gamed their internal AI usage leaderboard with low-value prompts, driving up infrastructure costs without meaningful output. They shut down the leaderboard.

Microsoft learned this when Claude Code costs hit $500 to $2,000 per engineer per month for heavy users. They revoked licenses broadly.

Uber learned this when AI tool adoption accelerated faster than the finance team modeled it. They burned through the annual budget in the first four months of the year.

Anthropic offers enterprise controls: admin dashboards, per-user spending caps, usage limits, compliance tooling. The features existed. They just were not turned on.

The lesson is not that AI tools are too expensive. The lesson is that uncapped AI usage at scale is a new category of financial risk that most companies have not built governance for yet.

What to Actually Do

If you pay for GitHub Copilot Pro and use it heavily, run the April usage preview report in your billing dashboard before your June bill arrives. GitHub made this available specifically to let users understand what their new costs look like before the first cycle closes.

If you manage a team on Copilot Business or Enterprise, set organization-level spending limits today. The controls are in Settings under Billing and Plans. Set a hard monthly cap and a notification threshold at 70% of it.

If you are evaluating alternatives, compare total cost of ownership across realistic usage scenarios, not just the headline monthly price. A $20/month flat-rate tool that you use heavily might cost the same as a $10/month credit-based tool at the same usage level. Run the numbers at your actual session length and frequency.

Diversify your AI tool stack across at least two platforms. Single-tool dependence creates both supply risk (if pricing changes dramatically) and cost risk (if one platform changes its model). Most developers who use Cursor for agentic sessions also use Claude.ai for reasoning tasks and keep a lighter Copilot plan for IDE completions.

For companies deploying AI at scale: implement spending caps before you deploy, not after. Treat your AI API budget the same way you treat your AWS spend. Set limits. Set alerts. Assign ownership.

The Honest Forecast

The unlimited, all-you-can-use AI era was a land-grab pricing strategy. It worked. Tens of millions of developers built habits around AI tools. Companies integrated AI into their core workflows. The dependency is real.

Now the platforms are transitioning to sustainable economics. GitHub moved first because the cost pressure from agentic workflows became impossible to absorb at a flat rate. Others will follow, on their own timeline, with their own framing.

This is not a reason to stop using AI tools. The productivity gains are real. The capability improvements continue. But the pricing mental model needs to update.

Flat-rate unlimited AI was a feature of 2022 and 2023. In 2026, the right question is not which tool is cheapest. It is which tool gives you the most predictable cost structure for your specific usage pattern. That is a harder question. It is also the right one.

Sources: GitHub official blog (github.blog), GitHub community discussion thread 192948, Axios report via Fast Company and Tech Startups (May 28, 2026), Visual Studio Magazine developer survey, GitHub Copilot plans page (github.com/features/copilot/plans).