“Unlimited Use for a Flat Monthly Fee”—That Premise Is Starting to Collapse
“Pay $20 a month and use the latest AI as much as you want.”
That dream-like era is quietly but surely coming to an end.
In April 2026, GitHub announced that GitHub Copilot will shift all plans to usage-based billing starting June 1, 2026. Around the same time, OpenAI’s ChatGPT lead openly stated that “unlimited plans are no more rational than unlimited electricity,” and Anthropic’s Claude is also reviewing its flat-rate plans.
This article examines why flat-rate plans are collapsing so rapidly right now, and what users need to know and do in response.
What Is Happening—Trends Across Major Services
GitHub Copilot: All Plans Move to Usage-Based Billing on June 1, 2026
On April 27, Microsoft-owned GitHub announced that all plans for its AI coding assistant “GitHub Copilot” will move to usage-based billing on June 1. Monthly fees will remain the same, but credits will be consumed based on actual usage, and users who exceed the credits bundled with their plan will need to purchase additional credits.
The backdrop is that the spread of agent-style coding has doubled weekly operational costs since the start of 2026, breaking the economics of the flat-rate, request-based model.
Specifically, from June 1, 2026, Copilot’s billing unit will transition to “GitHub AI Credits.” One credit equals $0.01, calculated based on the consumption of input tokens, output tokens, and cache tokens. Under the same act of “using Copilot,” costs will differ significantly between a short question and a long agent run.
Note that code completion and next edit suggestions will remain unlimited on paid plans. An important point: “not everything is moving to usage-based billing.”
OpenAI (ChatGPT): Hinting at the End of “Unlimited Use”
Nick Turley, head of ChatGPT at OpenAI, revealed plans to change the company’s AI pricing structure, hinting that “unlimited” subscriptions may eventually disappear. “Offering an unlimited-use plan in this era may be like offering unlimited electricity. It just doesn’t make sense,” he said.
OpenAI CEO Sam Altman has also stated that as AI demand surges, “AI may come to be sold on a usage-based model, billed by the amount consumed, just like electricity.”
Anthropic (Claude): A Redesign for the Age of Agents
On April 15, 2026, Anthropic announced that it will move the Claude Enterprise plan from a fixed-price model to a dynamic, usage-based billing model.
Amol Avasare, Head of Growth at Anthropic, officially explained: “When we launched the Max plan a year ago, Claude Code, Cowork, and agents that run for hours didn’t exist. Usage patterns have fundamentally changed, and the current plans aren’t built for this shape of work.”
Why Are Flat-Rate Plans Collapsing Now?
Cause 1: AI Agents Consume Resources on a Completely Different Scale
Previously, AI was fundamentally based on “the user asks, the AI answers.” But between 2025 and 2026, autonomous AI agents such as Claude Code and GitHub Copilot’s agent mode spread rapidly.
Every time an agent performs a chain-of-thought output, it consumes millions of tokens. Flat-rate subscription revenue simply cannot absorb this compute cost anymore—that is the honest truth from both companies.
It has become standard for a single user to run 4 to 10 agents in parallel. This is like “one buffet customer eating from 4 plates simultaneously.”
Cause 2: Flat-Rate Plans Were “Deficit-Ready” from the Start
The flat-rate plans from Anthropic, Google, and OpenAI were all designed as “loss leaders”—services run at a loss to acquire customers. Now that agent features have been opened up to individual-tier plans, the inference cost per user has far exceeded initial assumptions.
GitHub Copilot sells for $19/month, but the actual cost exceeds $60—a roughly $40/month loss per user. The more subscribers, the bigger the losses—just as a buffet owner weeps when more guests show up. That has been the reality of the cloud AI industry for the past two years.
Cause 3: Reasoning Models Cost More the More They “Think”
“Test-time scaling,” where AI becomes smarter by using more compute at inference time, is becoming mainstream. CoT (chain-of-thought) models have higher inference costs than traditional LLMs, creating a structure where the smarter the AI you use, the more the cost balloons.
Impact on Users—What Will Change
An Era Where You Can’t Predict Your Monthly Budget
The biggest change for users is that you can no longer predict at the start of the month “what you can do this month for this amount of money.” Token consumption varies wildly by model. For long agent sessions, cases of consuming millions of tokens per session have been reported, meaning just a few heavy sessions a month can exhaust your quota.
Corporate Costs Could Increase by 1.5x to 3x
For medium-to-heavy usage, costs are expected to rise to 1.6–2.6x current levels. Exchange rate fluctuations could push them even higher. For Japanese companies, “continuing current operations as-is” realistically means an annual budget increase of 1.5x to 3x.
Transition Support Measures for Corporate Customers
Transition support is available for corporate customers. Existing Business and Enterprise customers will receive more credits than usual during a three-month period from June to August. A mechanism to pool and share credits within an organization has also been introduced, allowing administrators to set budget caps at the company, department, or individual level.
What Should Individual Users Do?
Understand Your Own “Usage Pattern”
In the usage-based world, costs vary dramatically based on which model you use, how you use it, and how much you use it.
Start by organizing your own usage style:
- Light, chat-focused users: Flat-rate plans may still be advantageous in some cases
- Heavy users of agent coding: You’ll be directly impacted by usage-based billing
- Irregular users: Usage-based billing is more rational than flat-rate, which charges even in months you don’t use the service
A Strategy of “Using Multiple Services Appropriately”
A “two-tier system”—pure token-based billing for heavy users and flat rates for light users—is likely to become the realistic answer in the second half of 2026.
Rather than depending on a single service, using different tools for different purposes will be the key to cost optimization.
Make “Visibility of Consumption” a Habit
It’s essential to build a mechanism to measure token consumption. Before jumping to “budget cuts,” visualize who is using how many tokens on what. You can check this via the organization usage dashboard for GitHub Copilot, the usage command for Claude Code, and the organization dashboard for OpenAI.
The New Common Sense of AI Usage Going Forward
AI service pricing is now converging toward an infrastructure-style usage-based model, like electricity, water, or gas.
The era of “cheap and unlimited” is ending, but that’s not necessarily a bad thing. Paying only for what you use also means that with the right usage, you won’t pay wasteful costs.
What matters is shifting from hunting for “the cheapest plan” to “choosing the service and billing model that best fits your usage”.
AI has become an “infrastructure” to be used wisely with cost in mind. Those who embrace this mindset will ride the coming AI era skillfully.