Pay Up, Sucker.

This week sucked if you are a developer who built a product around Github Copilot's $20/month unlimited access. The company just called their own pricing model "broken, busted, and unsustainable" and cut the tier. The shock is real, but the adjustment was invisible. Nobody told you the cost was fiction, and you built on it.

Anthropic made the same admission with less drama. Claude Pro and Max tiers were designed for chat workflows. Not for agentic tasks, not for Claude Code running multi-step autonomous loops, not for anything that actually hammers the API. The tiers were priced for a use case that's already obsolete. The O'Reilly Radar for June 2026 flagged this as a structural shift across the industry. Metered billing is the new architecture.

To put it in perspective, I own a house built in 1925. It runs on oil heat.

When this house was built, heating oil was cheap enough that insulation was an afterthought. Efficiency has no value when energy has no price.

Then the 1970s happened. The crisis didn't change the laws of thermodynamics, but it made the bad decisions visible and painful.

The $20 AI subscription was the same bet. The AI companies subsidized compute to build dependency, to lock in users, to grow market share fast enough to matter. "The first one's free" is a classic pusher strategy. It worked. Teams built workflows, integrated APIs, wrote internal tools that assumed unlimited cheap inference. The insulation problem was invisible because the energy was free.

Now it isn't. This was always the plan.

Low cost usage does unlock experimentation. In the discovery phase of any new technology, removing cost from the equation isn't irrational. You need to try things. Pirated copies of Photoshop back in the day made more designers who were dependent on Photoshop, which ultimately made Adobe more money. The foundation model companies understood this; subsidized pricing bought them exactly what they wanted: a generation of products and workflows that now depend on their APIs.

The experimental phase is over. The bill is how you know it. AI isn't going anywhere — let that go — but an adjustment is coming for everyone who treated the discovery phase like a permanent business model. The developers and companies that get ahead will be the ones who know how to be self-sufficient.

The teams that are already fine are the ones who treated AI compute as a real cost before anyone made them. Developers replacing entire browser extension stacks with local LLMs didn't do it because they saw this coming. They did it because they got tired of paying for things they could run themselves. Open-source tools doing what Claude charges for, sometimes better. Local design tools replacing Claude Design for teams who couldn't justify the subscription. And these aren't edge cases from frugal hobbyists. They're the leading edge of a cost-conscious shift that the pricing correction has just made impossible to avoid.

The infrastructure for this shift already exists. The NVIDIA DGX Spark runs models up to 200B parameters locally for under $3,000. Krea 2, ranked first among independent image generation labs, runs on your own hardware. The argument against local models used to be capability. That argument is getting harder to make every month.

The conventional wisdom says use the best frontier model for everything. That was reasonable when everything cost $20/month. The actual discipline, the one that survives a pricing correction, is using the cheapest model that works and reserving frontier API calls for tasks that genuinely need them. Summarization, text cleanup, basic generation, classification: none of that needs Claude Fable 5. Running it through Fable 5 anyway is the AI equivalent of leaving the windows open in January because the furnace is cheap.

Most teams don't have this discipline yet because they've never needed it. That's going to change fast.

The AI companies that subsidized compute to create lock-in have now created the exact conditions for their own disruption. By making API calls expensive, they've made open-source and local model adoption economically rational for the first time. The teams most likely to thrive are the ones most willing to leave.

I'm still working on my 1925 house. We spent $25,000 on spray foam insulation last fall. I like restoring things, making things better, but it's less than ideal. Sadly, the longer you wait to make things efficent, the more costly it will be in the long run.

The retrofit begins now.