A follow-up to The End of the AI Gold Rush
In August 2025, I wrote that the AI industry’s $20-a-month illusion couldn’t last — that the math simply didn’t work, and that eventually someone would expect their profit out of this. I estimated that reaching actual cost recovery would require 10x to 50x price corrections. I did not expect the opening moves to arrive within a year, or to arrive simultaneously from two of the biggest names in developer tooling.
This week, both Anthropic and GitHub made moves that confirm the subsidy cliff is no longer theoretical.
Anthropic Tests the Edges
On April 21st, Anthropic briefly removed Claude Code from the $20/month Pro plan for new subscribers. Ed Zitron at Where’s Your Ed At documented what happened: support documentation was quietly updated to say “Using Claude Code with your Max plan” instead of “with your Pro or Max plan,” and pricing pages across both mobile and desktop stopped showing Claude Code as a Pro benefit.
Anthropic’s Head of Growth later characterized it as a small test affecting roughly 2% of new signups, and the changes were quietly reversed. But the support documents and pricing pages don’t update themselves by accident. Nor does a company that wants to keep something quiet allow their support docs to contradict their marketing copy at the same time on both desktop and mobile.
What’s actually going on is not hard to read: Anthropic’s Head of Growth acknowledged that Claude Code and Claude Cowork were released after the Max plan was designed, and that the Max plan was originally built “for heavy chat usage, that’s it.” In other words, the current pricing tiers weren’t designed with agentic, multi-hour coding sessions in mind. They were designed for a product that no longer exists.
GitHub Makes It Official
While Anthropic was testing the water quietly, GitHub was more direct. Starting June 1st, all GitHub Copilot plans will move to usage-based billing via a new “GitHub AI Credits” system. The math is straightforward: each plan’s monthly price converts directly into credits of the same dollar value. Pro at $10/month gets $10 in credits. Business at $19/user/month gets $19 in credits.
The rationale in GitHub’s announcement is essentially the same argument I made in August: agentic usage has fundamentally changed the economics. A quick chat question and a multi-hour autonomous coding session don’t cost the same to run, so it’s no longer sustainable to charge users the same amount for both.
There’s a notable detail buried in the announcement: code completions and Next Edit suggestions are explicitly excluded from credit consumption. Those stay free. What gets metered is everything else — the multi-turn conversations, the repository analysis, the agentic sessions. The high-value, high-compute work. The part that has been effectively subsidized.
For business and enterprise customers, GitHub is providing a promotional grace period through August — higher included credits than their subscription price would normally buy. That’s not generosity; that’s an onboarding ramp designed to prevent sticker shock from derailing adoption.
The Pattern Is Clear
These two moves, arriving within a week of each other, aren’t coincidental. The entire AI developer tooling market is working through the same problem at the same time: agentic usage is the future everyone wanted, and it costs dramatically more than the subscription models that got us here.
In August I wrote about a theoretical market stratification — an AI Elite with full access, an AI Middle Class operating under token limits, and AI Have-Nots returning to pre-AI methods. That framing was speculative then. It’s less speculative now. GitHub’s model makes the tiers explicit: you have included credits, optional additional credits at published API rates, and admin-level budget controls that can cut you off entirely when the pool is exhausted.
The fallback experience is gone. Previously, Copilot users who burned through their premium requests could fall back to a lower-cost model. Under the new system, you either have credits or you stop. The floor has been removed.
What Hasn’t Changed (Yet)
Base plan prices aren’t changing. $10 is still $10, $19 is still $19. What is changing is what those prices actually buy in terms of heavy usage. For light users, this transition is nearly invisible. For power users who’ve been treating these tools as unlimited resources, June 1st is a reckoning.
The Anthropic situation is still unresolved. Claude Code appears to still be available on Pro plans for existing subscribers, but the signal from that April 21st test is that the current pricing doesn’t work for what Claude Code has become. Whether the correction comes as a tier restriction, a usage cap, or a separate pricing model, something is changing.
What I’d Revise From August
My August post assumed the correction would be more dramatic — a sudden cliff edge. What’s actually happening looks more like a managed descent: usage-based billing, promotional grace periods, grandfathered annual plans, admin budget controls. The companies doing this have learned from watching what happened to Heroku and other platforms that tried to abruptly normalize pricing after years of subsidy. They’re trying not to detonate their user bases.
That doesn’t change the underlying math. It just means the squeeze arrives gradually rather than all at once.
The other thing I’d update: the beneficiaries of this shift may not be experienced engineers quite as quickly as I imagined. What we’re seeing first is a squeeze on individual developers who’ve been using these tools in ways their subscription plans were never priced to support. The enterprise market is actually being handled more carefully — higher promotional credits, pooled usage, budget controls. The transition is being absorbed at the organizational level before it filters down to individual cost pressure.
But the direction is the same. The free buffet is closing. What replaces it will reflect actual costs, and actual costs are not $20 a month for a multi-hour agentic coding session on a frontier model.
The Bottom Line
I wrote in August that the gold rush was ending. This week, two of the biggest players in developer AI started publishing the closing times. The Subsidy Cliff is no longer a prediction — it’s a schedule.
For developers, the practical implication is that the way you use these tools needs to evolve. Knowing when to invoke an agentic session versus a quick targeted query is becoming a real efficiency skill. Prompt crafting that minimizes token waste is no longer just a power-user optimization; it’s cost management. And the question of which tasks actually justify the compute cost — versus what you can do faster and cheaper without AI — is going to matter more than it did when the meter wasn’t running.
The Easter Bunny and Santa Claus are still in the room. But they’re clearly beginning to check their watches.