There is a version of the AI industry story that positions every pricing change as a sign of maturity. Costs come down, products improve, value compounds. That is the story the industry has told, with remarkable consistency, since the first wave of generative AI subscriptions landed in 2022. It is also, according to writer and analyst Ed Zitron, largely fiction.
The evidence for that claim just got harder to ignore. On April 27, 2026, GitHub Copilot confirmed that all its plans would move to usage-based pricing on June 1, 2026. Users who previously paid a flat monthly fee would now be billed based on the actual cost of the models they use. The move was framed by GitHub as an alignment of pricing with real usage. Zitron's reading is more direct: Microsoft could no longer afford to keep subsidizing nearly two million people's compute.
The Observation
Zitron's newsletter essay, published April 28, 2026, argues that the monthly subscription model has never made economic sense for any service connected to a large language model. His core position is that AI companies deliberately obscured the true cost of inference in order to grow user bases, creating a dependency before presenting customers with the actual bill. The GitHub Copilot announcement, he writes, is the most visible crack in a structure that was always going to fracture.
The Numbers Were Never Hidden
The Wall Street Journal reported in October 2023 that Microsoft was losing on average more than $20 a month per GitHub Copilot user, on a plan that cost individuals $10 a month. Some users were costing the company as much as $80 a month. That gap did not close. It widened. Zitron notes that Anthropic, until recently, allowed users to burn upwards of $8 in compute for every dollar of their subscription. One user on the GitHub Copilot subreddit calculated that the token burn of what used to be a single premium request ran to roughly $11, because one request could involve 60,000 tokens in the context window, multiple tools, and numerous internal turns.
The Gym Membership That Was Never a Gym Membership
Zitron uses a blunt structural argument to explain why flat subscriptions work for some businesses and cannot work for AI. A gym can model wear, staffing, and electricity costs with reasonable predictability. A Google Workspace user, before AI, cost whatever digital storage and low-compute document services amount to. Neither varies wildly by user behavior. An AI subscriber is entirely different: one person might use ChatGPT for a weekly search while another feeds in entire codebases. The provider has no lever to pull except degrading the product, shrinking context windows, or pushing users toward cheaper models. Every one of those levers has been pulled.
The Deliberate Fog Around Token Costs
Zitron's sharpest charge is that the opacity was intentional. AI companies used tokens, messages, percentage gauges, and five-hour rate limits to prevent users from ever seeing a dollar figure attached to a task. Anthropic's Claude Code developer documentation stated, until the beginning of April 2026, that the average cost was $6 per developer per day, with 90% of users staying below $12 a day. That language has since been revised. Anthropic Head of Growth Amol Avasare stated publicly that Max subscriptions were built for heavy chat usage rather than what users are doing with Claude Code, and signaled that Anthropic is exploring different options to keep delivering a great experience. Zitron translates that as an announcement that prices will change.
Enterprise Exposure Is Already Breaking Through
The cost problem is not limited to individual subscribers. Uber's CTO said at a conference that the company had spent its entire AI budget for 2026 in a matter of months. Goldman Sachs has suggested some companies are spending as much as 10% of their headcount costs on AI tokens, with projections that the figure could reach 100% within a few quarters. A 10-person Anthropic Teams subscription at $1,250 a month, Zitron estimates, likely burns between $5,000 and $10,000 a month in actual API calls. Both Anthropic and OpenAI have already moved enterprise customers to token-based billing. The consumer tier is next.
Microsoft as the Canary
Zitron flags the GitHub Copilot move as significant beyond its immediate user base. Microsoft is, by his assessment, the best-capitalized and best-positioned company in the AI infrastructure space. If Microsoft cannot sustain subsidized compute for its developer tool, no company further down the capitalization ladder can either. GitHub Copilot's previous pricing allowed 300 premium requests a month, with unlimited chat using models including GPT-4o mini. Microsoft gave users completely unlimited model access until May 2025. Each step toward restriction has been met with user backlash, and the subreddit response to the June 1 announcement follows that pattern, with users describing the product as dead.
Zitron's argument lands against a backdrop where every major study on AI's return on investment struggles to find concrete evidence that it exists. The subscriptions made the economics feel acceptable. Once the actual per-token costs surface, the question of whether any of this spending is justified becomes much harder to wave away.
If the trajectory Zitron describes continues, the signal to watch for is a major lab such as Anthropic or OpenAI moving all consumer subscribers to token-based billing. Should that happen, the flat-fee era of generative AI could close entirely, and the creative and marketing industries that built workflows around $20-a-month access may find the math of those workflows no longer holds. The dependency was the product. The price increase was always the plan.