Rationing: The Industry

On April 20, 2026, GitHub paused new signups for Copilot Pro, Pro+, and Student plans. On April 27, GitHub announced that all Copilot plans would move to token-based billing on June 1. On May 6, Anthropic doubled Claude Code's rate limits after securing a compute deal with SpaceX. Between those dates, Cursor overhauled its pricing for the second time in a year. Every major AI coding tool repriced within the same six-week window. That is not a coincidence. That is an industry discovering it cannot afford its own product.

We have written thirteen articles in this series. The Rationing documented the pattern at Anthropic. Pay-Per-Overage tracked the billing shift. The Cache exposed the prompt caching changes. Claude Code measured the degradation. Copilot Credits caught GitHub's turn. The Double Cross showed that even Anthropic's "fix" came with unchanged weekly caps.

Every one of those pieces focused on individual actors making individual decisions. This piece is different. The pattern has gone systemic. When every vendor in a market reprices simultaneously, that tells you something about the market's structure, not just any single company's finances.

The Subsidy Era Is Over

In October 2024, the Wall Street Journal reported that Microsoft was losing more than $20 per user per month on GitHub Copilot, which was priced at $10. Some heavy users were costing Microsoft as much as $80 per month in compute. The company was subsidizing every seat, burning cash to build market share in AI-assisted development.¹

By January 2026, the subsidy had grown untenable. According to documents obtained by Ed Zitron at Where's Your Ed At, the weekly cost of running GitHub Copilot had nearly doubled since the start of the year. The cause: agentic workflows. A single multi-hour coding agent session can consume more tokens than an entire month of traditional code completions. The fixed subscription price that once covered a predictable number of autocomplete suggestions now had to absorb open-ended, parallelized agent sessions that ran for hours.²

GitHub's CPO Mario Rodriguez said the quiet part out loud: "The current premium request model is no longer sustainable."³ On the Q3 2026 earnings call, Satya Nadella confirmed that Microsoft's productivity and coding tools would transition to "per user and usage" pricing. The flat-rate era was officially dead.

This is not just a Microsoft problem. It is the fundamental economics of inference. Running a frontier model costs real money per token. The more capable the model, the more it costs. The more agentic the workflow, the more tokens it burns. And every AI coding vendor set their initial prices before agentic workflows existed, when usage meant short autocomplete suggestions, not autonomous multi-file refactors that run unsupervised for twenty minutes.

The June 1 Cliff

On June 1, 2026, every GitHub Copilot plan switches from request-based billing to token-based billing using "AI Credits," where 1 credit = $0.01 USD. The math is straightforward and worth spelling out.⁴

Copilot Pro costs $10 per month. It now includes $10 in monthly AI Credits. Copilot Pro+ costs $39 per month. It includes $39 in credits. Copilot Business: $19 per user per month, $19 in credits. Copilot Enterprise: $39 per user per month, $39 in credits. Code completions and Next Edit suggestions remain unlimited. Everything else, every chat message, every agentic session, every code review, consumes credits based on actual token usage at published API rates for each model.

One developer did the math on their actual Pro+ usage. Under the new system, their last month would have cost $52 instead of $39, a 33 percent increase for an identical workflow.⁵ That is the optimistic scenario. Developers who rely on Opus 4.7 for complex agentic tasks will burn through $39 in credits far faster than those using cheaper models. The multipliers are steep: Claude Opus 4.7 carried a 7.5x request multiplier before the switch. In credit terms, a single Opus conversation costs roughly seven times what the same conversation costs with a lighter model.²

The headline price did not change. The amount of work that price buys did. Visual Studio Magazine captured the developer reaction accurately: "You will get less, but pay the same price."⁵

The Sequence

What happened at GitHub did not happen in isolation. Here is the timeline:

April 20: GitHub pauses new signups for Copilot Pro, Pro+, and Student plans. Removes Opus models from the $10 Pro tier. Tightens usage limits. States that agentic workflows have "fundamentally changed Copilot's compute demands."⁶

April 22: GitHub pauses new self-serve signups for Copilot Business on Free and Team plans.⁷

April 27: GitHub announces all plans move to usage-based billing on June 1.⁴

May 6: Anthropic announces the SpaceX compute deal. Doubles Claude Code rate limits. Removes peak-hour caps for Pro and Max. The messaging is about capacity, but the context is thirteen weeks of rationing complaints, five-hour usage windows, and weekly hard caps that never moved even as the company promised improvements.⁸

Concurrent: Cursor, which overhauled its pricing in June 2025 by replacing fixed "fast requests" with usage-based credit pools, continues adjusting. The Pro plan now includes $20 in monthly usage for frontier models. Usage through Cursor's auto-routing is unlimited, but premium model access draws from the pool. Teams pricing followed the same pattern in August 2025.⁹

Three companies. Six weeks. All moving in the same direction: from flat pricing to metered consumption. From "unlimited" to "it depends on how much you use."

The Playbook

Cory Doctorow named this pattern in 2022. He called it enshittification, and described its three stages: first, the platform subsidizes users to build market share; then, it subsidizes business customers to build an ecosystem; finally, it extracts maximum value from both.¹⁰

The AI coding tool market has compressed all three stages into roughly eighteen months. GitHub Copilot launched at $10 per month in June 2022, losing $20 per user from day one. Cursor offered 500 premium requests per month at $20. Anthropic priced Claude Code into the $20 Pro plan. Each vendor set prices that had nothing to do with costs and everything to do with adoption curves.

The extraction phase arrived faster than anyone expected, because agentic workflows broke the math. A developer using Copilot for autocomplete in 2023 consumed a predictable, manageable amount of compute. A developer running Copilot's coding agent for multi-file refactors in 2026 consumes orders of magnitude more. Zitron's reporting put numbers on it: weekly infrastructure costs nearly doubled in four months.² The subsidy that was already bleeding cash at 2023 usage levels became catastrophic at 2026 usage levels.

This is not a new pattern in technology. Uber subsidized rides with venture capital, then implemented surge pricing once the taxi industry was gutted. Facebook gave publishers enormous free reach, then throttled organic distribution and sold the reach back as ads. Amazon Web Services priced below cost to kill competitors, then raised prices methodically once customers were locked in.¹⁰ The mechanism is always the same: make it free until it's essential, then make it expensive because it's essential.

The AI coding tool industry followed this playbook at startup speed. The difference is that every vendor hit the wall at the same time, because they all depend on the same underlying resource: GPU inference from the same handful of model providers.

The Compute Ceiling

The simultaneous repricing is not coordination. It is collision with a shared physical constraint. Every AI coding tool runs on a small number of frontier models, all of which require massive GPU clusters to operate. When demand outpaces supply, every product built on that supply degrades at once.

Anthropic acknowledged this directly. The company told Semafor it had planned for 10x growth in Q1 2026 but experienced 80x.¹¹ That gap between projected and actual demand explains every rationing decision we have documented in this series: the five-hour windows, the weekly caps, the peak-hour throttling, the model downgrades. Anthropic was not being greedy. It was being overwhelmed.

The SpaceX deal brings over 300 megawatts of new capacity, more than 220,000 Nvidia GPUs, from SpaceX's Colossus 1 data center.⁸ Goldman Sachs projects $7.6 trillion in total capital deployment across compute, data centers, and power infrastructure between 2026 and 2031.¹¹ Google processes 16 billion tokens per minute, up 60 percent quarter-over-quarter, while remaining compute-constrained. AWS allocated $43.2 billion to Q1 capital expenditure alone.³

These numbers represent the largest infrastructure buildout since the electrification of cities. And they are not enough. Demand continues to outpace supply because the product itself generates more demand: the better AI coding tools get, the more people use them, and the more they use them, the more tokens they consume. Copilot grew from 15 million to 20 million paid enterprise seats in a single quarter. Weekly engagement now matches Outlook.³ Each of those 20 million users is consuming more tokens per session than they were three months ago, because each update makes the tools more capable and more agentic.

You cannot flatten a demand curve that your own product steepens.

The Price of Alternatives

While the incumbents raise prices, the floor is falling out beneath them.

DeepSeek V4 Flash offers inference at $0.14 per million input tokens and $0.28 per million output tokens. DeepSeek V4 Pro, a model competitive with frontier offerings on coding benchmarks, costs $1.74 per million input tokens at list price, currently discounted 75 percent to $0.44 through the end of May. For cached input, the price drops to fractions of a cent.¹²

Compare those numbers to what GitHub Copilot charges. Claude Opus 4.7 through Copilot's credit system costs many times more per token than running the same model through Anthropic's API directly. The credit system adds a layer of abstraction that obscures the actual token cost, which is the point. If developers could easily compare what Copilot charges per token to what the raw API charges, the markup would be difficult to justify.

We covered this dynamic in The Arbitrage Agent. Developers are building tools that route queries to the cheapest capable model rather than the most expensive default. Projects like DeepClaude use frontier models for reasoning and cheaper models for generation, cutting costs by 60 to 80 percent while maintaining output quality.¹³ The arbitrage opportunity exists because the incumbents are not pricing to cost. They are pricing to recoup years of subsidies.

The market is splitting into two tiers. The top tier charges premium rates for integrated experiences: IDE plugins, agentic workflows, code review, everything in one subscription. The bottom tier offers raw inference at commodity prices. The gap between those tiers has widened from roughly 2x to 10x or more in the past six months. That gap is the subsidy recovery tax.

What the Contracts Say

In March, Nadia Byer reviewed the terms of service for every major AI coding tool and published her findings in The Rationing: What the Contracts Actually Say. Every vendor reserves the right to change pricing, throttle usage, modify model availability, and alter service levels at any time, with minimal notice.

Those contractual escape hatches have now been exercised simultaneously. GitHub removed Opus from Pro and paused signups with one day of notice. Anthropic changed its caching behavior without announcing it at all, as we documented in The Cache. Cursor's mid-2025 pricing overhaul generated enough unexpected charges that the company had to issue a public apology and offer refunds.⁹

The contracts are not bugs. They are the mechanism by which the extraction phase operates. Every AI coding tool vendor wrote terms that let them change the deal whenever the economics demanded it. The economics now demand it.

The Anthropic Paradox

The SpaceX deal illustrates the contradiction at the heart of this industry. Anthropic doubled Claude Code's rate limits on May 6 and announced the removal of peak-hour usage caps.⁸ The company framed this as giving users more. But the weekly caps that triggered the most user complaints, the ones we documented in The Double Cross, remained unchanged.

The doubling is real. Users can now consume twice as much within a five-hour window. But if the weekly ceiling stays fixed, a higher hourly rate just means you hit the wall faster. It is like doubling the speed limit on a road that still has the same toll booth at mile 50.

Anthropic's revenue grew 80x in one quarter. If that growth continues, the 300 megawatts from SpaceX will be consumed as fast as it comes online. The rationing did not end. The rationing changed shape. Instead of throttling everyone equally, the new model throttles by price tier: Max 20x users at $200 per month get the largest allocations, followed by Max 5x at $100, and Pro users at $20 get whatever remains. The scarcity is being managed through pricing rather than through blunt rate limits, which is more elegant but still rationing.

The Lock-In Problem

A developer who has spent six months building workflows around Copilot's agentic features, configuring context files, training muscle memory on keyboard shortcuts, integrating with GitHub's PR review pipeline, does not casually switch to a competitor when the price goes up 33 percent. The switching cost is real, measured in hours of reconfiguration and weeks of reduced productivity during the transition.

This is the leverage point the subsidy strategy was designed to create. Four years of subsidized pricing built deep integration into millions of developers' daily workflows. The June 1 billing change arrives after that integration is complete. The timing is not accidental.

The same lock-in applies to Anthropic's Max plans. Developers who built their tooling around Claude Code's specific capabilities, its tool-use patterns, its file-editing conventions, its particular strengths in refactoring and testing, face real migration costs. The $200 monthly price tag for Max 20x is justifiable only if the alternative is rebuilding your entire workflow from scratch. And for many developers, it is.

Copilot has 20 million paid enterprise seats.³ Each seat represents a developer who has integrated AI coding into their daily practice, often at their employer's direction. When the price goes up, most of those developers will absorb it, because their employer decided the tool was mandatory and the switching cost is too high to reconsider. That is the point of the subsidy phase. You do not subsidize to be generous. You subsidize to make the extraction phase possible.

What Comes Next

The industry is not going back to flat pricing. The economics do not support it. Every company that offered unlimited AI coding at a fixed monthly rate has either already moved to usage-based billing (Cursor, GitHub) or is managing demand through rate limits and tier restrictions (Anthropic). The prediction from one industry analysis that Cursor, Windsurf, and "every other AI coding tool" would follow GitHub's billing model within 60 to 90 days looks increasingly conservative.¹⁴

The transition will happen in stages, as it already has. First, remove the most expensive models from the cheapest tiers. Then tighten usage limits. Then pause new signups to stop the bleeding. Then announce "usage-based billing" as though it were an upgrade rather than a price increase. Then offer a promotional credit buffer during the transition to soften the impact. Then let the buffer expire and let the real prices take hold.

GitHub is executing this sequence in public, in real time. Anthropic executed it over the course of thirteen weeks, which we documented in thirteen articles. Cursor executed it with less public drama but the same structural outcome. The sequence is the same because the underlying constraint is the same: inference costs money, demand is growing faster than capacity, and the subsidies were never meant to be permanent.

The honest framing would be: "We priced this product below cost to acquire market share. Now we need to charge what it actually costs." No vendor will say that. Instead they say "agentic workflows have fundamentally changed compute demands," as though the compute demands are the users' fault rather than the direct result of the product working as advertised.

The Structural Truth

This series began on March 27 with a simple observation: Anthropic was giving users less while charging the same. Thirteen articles later, the observation has scaled. The entire AI coding tool industry is giving users less while charging the same, or more, and doing it simultaneously.

The explanation that fits all the data is not greed, incompetence, or conspiracy. It is structural. Frontier AI inference is expensive. Subsidized pricing attracted demand that the infrastructure cannot serve at the subsidized price. The subsidy is ending because it has to end.

What makes this worth documenting is not that it is surprising. Subsidize-then-extract is one of the oldest plays in platform economics. What makes it worth documenting is that an entire industry ran the play at the same time, on the same product category, affecting the same users. A developer in May 2026 faces usage-based billing from GitHub, credit pools from Cursor, tiered rate limits from Anthropic, and a DeepSeek alternative that costs 1/107th the price of Claude Opus at list rates. The market is repricing in real time, and the developers who built their workflows during the subsidy era are the ones absorbing the cost.

The tools are still useful. Many developers will pay the higher prices and keep working. But the promise that sold the first 20 million seats, unlimited AI coding for a flat monthly fee, is gone. It was never a real promise. It was a customer acquisition cost, amortized across millions of users, funded by investors who expected the extraction phase to arrive eventually.

It arrived.