Three separate stories this week, from three different corners of the AI world, all point at the same shift. The model layer, the thing everyone treated as the scarce and valuable part, is turning into a commodity. Cheaper, more interchangeable, and harder to build a business on by itself.
Read alone, each is a product update. Read together, they are a pricing signal.
Signal One: A New Frontier Model, on Schedule
OpenAI previewed its next-generation model this month, GPT-5.6 Sol. The capability story matters less than the cadence. Frontier models now arrive on a predictable rhythm, each one resetting the ceiling a few months after the last.
For a buyer, predictable improvement changes the calculus. When the best model is a moving target that upgrades every quarter, paying a premium to lock into today's version makes little sense. You are renting a capability that will be standard, and cheaper, by the time you finish integrating it. Abundance on a schedule is the opposite of scarcity, and scarcity is what lets a vendor hold a price. This is the source crediting OpenAI's own preview, and the cadence is the real message inside it.
Signal Two: Frontier AI at Half Price
Anthropic struck a deal with California's governor to give state agencies access to Claude at half price, with training and support attached. TechCrunch framed it as cost management for a public buyer, which it is. It is also a number, and numbers are honest.
When a frontier lab cuts its price in half to win a large account, the signal is that the model is no longer the rare thing it can charge a premium for. Competition and capable alternatives have arrived, and price is now a lever the labs are willing to pull. Enterprise AI costs got high enough that big buyers started pushing back, and the labs are blinking first.
I wrote in the last edition that your AI vendor can be switched off. The other side of that coin is that a vendor competing on price is a vendor that knows you can switch. Half-price Claude is what a buyer's market looks like when the buyer finally has options.
Signal Three: An Open Model Replacing the Premium One
The third signal is the loudest, because it came from practitioners rather than a press release. Builders are publicly swapping a cheaper open model, GLM 5.2, in for a premium one in their daily coding work, with the headline that they are replacing the expensive default. Lenny's Newsletter covered the switch as a real workflow decision, not a benchmark curiosity.
When the people who use these tools all day decide the cheaper option is good enough for the job, the premium option's pricing power erodes from the bottom. Good enough at a fraction of the cost is how commoditization actually happens. Not all at once, but one practical swap at a time, until the default quietly changes.
This is the pattern that should worry any company pricing AI as a premium product. The frontier labs are not being undercut by another frontier lab. They are being undercut from below, by open models that were unthinkable two years ago and now sit merely a notch behind. A notch behind, at a fraction of the price, wins more real workloads than the marketing wants to admit.
What to Build When the Engine Is Cheap
Put the three together. New frontier models on a schedule, premium models cutting price, and open models good enough to replace them in real work. The direction is one way. The model is getting cheaper and more interchangeable, fast.
If your strategy assumed the model was the moat, this is the week to drop that assumption. The defensible value is moving to everything around the model. Your proprietary data. The distribution you have earned. The judgment to use the output well and catch it when it is wrong. I argued the case for vendor optionality a few editions ago, and the smart posture has only sharpened. Stay portable. Assume the model underneath you will be swapped, repriced, or upgraded, possibly all three this year.
None of this means the frontier stops mattering. For the hardest problems, the best model still earns its premium, and it is worth paying for. But the share of work that genuinely needs the very best model is shrinking, and most companies were paying frontier prices for commodity tasks anyway. Match the model to the job and the bill drops without the output changing at all.
The companies that win the next phase will treat models the way they treat cloud compute. A necessary input, bought from whoever offers the best deal this quarter, never the thing the business is built to defend. The scarce resource was never the model. It was knowing what to do with it. That has not changed, and now it is the only part worth paying a premium for.