GitHub Copilot is abandoning flat-rate subscriptions. Starting June 1, 2026, Microsoft will charge users based on actual model costs rather than giving them a set number of requests. Microsoft has been bleeding money. The Wall Street Journal reported in 2023 that the company was losing over $20 per user per month on average, with some power users costing up to $80 monthly on a $10 subscription. GitHub's announcement admits that agentic usage, meaning multi-step autonomous coding sessions, drives much higher compute demands that made the old pricing model unsustainable.

Ed Zitron, who predicted a "subprime AI crisis" in 2024, argues the entire monthly subscription model for AI is fundamentally broken. Companies like OpenAI, Anthropic, and Microsoft subsidized compute costs to hook users, assuming token prices would drop over time. Instead, reasoning models and agentic tools burn more tokens than ever, pushing inference costs higher. Anthropic reportedly let users burn $8 in compute for every dollar of subscription revenue. Salesforce's Agentforce charges $2 per conversation. The math doesn't work when a flat monthly fee has to cover wildly variable compute costs. the $20 pricing era has an expiration date.

The picture is more complicated than Zitron suggests. Commenters on Hacker News, including people claiming industry experience, argue that frontier labs like OpenAI and Anthropic maintain healthy margins on token pricing, estimated at 80% or more. They aren't subsidizing users so much as running a standard subscription model where light users cover heavy ones. The real squeeze hits wrapper services like Cursor and Perplexity, which pay API costs at markup and resell to users. These companies sit in an uncomfortable middle ground between first-party providers with their own models and infrastructure, and open-source alternatives like Meta's Llama and Mistral's models that organizations can self-host.

Agentic AI breaks the model. Simple chat interactions have predictable costs. Autonomous agents running open-ended multi-step workflows? Those costs scale unpredictably. GitHub Copilot's move to usage-based pricing is a warning sign. The economics of AI agents, with their open-ended compute demands, don't fit the subscription mold that worked for simpler tools. For anyone building or buying agent products, pricing changes are coming. The only question is how fast.