For the past two years, choosing an AI model felt like the most important decision a company could make about artificial intelligence. Should it be ChatGPT, Copilot, Gemini, or Claude? That question is losing its importance fast, and the numbers explain why.
Stanford's AI Index Report tracked the cost of running an AI model at a fixed quality level. In November 2022, it cost twenty dollars to process a million words of text. By October 2024, that cost had fallen to seven cents. That is not a small price drop, it is a collapse of more than two hundred and eighty times in under two years, driven mostly by competition and cheaper computer chips.
DeepSeek, the Chinese AI company, pushed this further by pricing its models far below what OpenAI, Google, and Anthropic were charging, and still performing close to their level on most tasks. That forced the big American AI companies to lower prices too, since a model that costs ten times more but only performs marginally better is a hard sell to any finance department watching the bill.
This price war is why many companies now split their AI work across several models instead of picking one. A project from University of California, Berkeley called RouteLLM builds software that sends easy questions to cheap, fast models and only sends hard questions to expensive, powerful ones. Berkeley's own testing showed this kind of routing can cut costs by more than eighty percent while keeping almost all the answer quality of the expensive model.
Microsoft has read this shift correctly, and it explains why the company keeps investing in smaller, cheaper models under its Phi lineup, rather than only chasing the biggest model it can build. A smaller model trained on a company's own repair manuals, contracts, or customer records can outperform a giant general model on that narrow task, at a fraction of the running cost.
Here is the part that actually matters for any business leader. If the model is becoming cheap and swappable, then the company's advantage has to live somewhere else. Microsoft CEO Satya Nadella has been describing this directly, telling businesses to think in terms of two kinds of capital: the knowledge and judgment of their own people, and what he calls their own AI learning history, built from real decisions, results, and corrections over time. Anthropic has made a similar point in its own writing, arguing that companies get more out of managing the information and instructions surrounding an AI model than out of endlessly tweaking the model itself.
This has a very practical test attached to it. Ask what would happen to your company tomorrow if you switched from one AI provider to another. If most of your company's accumulated decisions, corrections, and outcomes are locked inside that one vendor's system and would be lost or hard to move, then that vendor, not you, holds the real value. If instead your own records of what worked and what did not live in systems you control, the model becomes a replaceable part, and switching costs you very little.
That is the shift worth planning around. Not which model to buy, but whether your company is building a memory of its own, or renting somebody else's.