What Anthropic’s pricing “test” reveals about the real cost of not owning your stack.
You do not own your AI. You rent it.
Last week, Anthropic proved it. They pulled Claude Code, their most popular developer tool from the $20/month plan and locked it behind a $100+/month tier. The backlash was so severe they reversed it almost instantly.
Most people moved on. That is a mistake.
This was not a pricing error. It was a preview of what happens when AI subsidies run out.
Every major AI subscription today is sold below cost. The labs know their pricing is unsustainable. They are buying market share now, planning to collect later.
When the correction comes, every workflow built on these tools becomes a cost you did not budget for, or a capability you lose overnight.
The natural reaction is to run models locally. Download the weights, host them on your own infrastructure, cut the dependency on someone else's pricing page.
That is a real step forward. Your data stays on your machines. You control uptime. No one can revoke access with a billing change.
But local does not mean independent.
Most downloadable models ship under licenses that restrict commercial use, limit modification, or let the vendor change terms. You solved the rental problem and inherited a licensing one.
In late 2024, the Open Source Initiative published OSAID 1.0, the first formal definition of Open Source AI. It requires the training code, the weights, and full transparency on the training data. Most models that call themselves "open" do not meet it.
Also last week, Moonshot released Kimi 2.6, an open-weight model closing the gap to frontier performance fast. You can self-host it, which helps with data privacy. But open weight is not open source. You cannot fully audit it, reproduce it, or guarantee independence from the vendor's future licensing decisions.
Better than a proprietary API? Absolutely. But it is a longer leash, not freedom.
True open source models do exist. Switzerland's Apertus ships under Apache 2.0 with full weights, training code, and data transparency. No licensing traps. Full auditability.
The technology is not the bottleneck. Most organizations are not even asking the right questions yet.
No framework to evaluate whether a license protects or traps them. No criteria to distinguish open source from open weight. No mandate to treat AI as infrastructure they govern, not a subscription they manage through procurement.
Anthropic's reversal was a free warning. The question is not whether you noticed, but whether you have the structure to respond.
If your AI strategy lives entirely inside a vendor's pricing page, it is not a strategy. It is a dependency.
