The price of running an AI task has fallen roughly 50 times in three years. In late 2022, processing one million words of text through a GPT-4-quality model cost around $20. By the end of 2025, the same quality cost $0.40. That is not a rounding error. It is a structural change in what AI costs to use at scale.
The force behind that collapse is competition, and most of the competitive pressure came from open-source models that anyone can download and run. When a comparable model is free to download, a paid service cannot charge $20 for long.
On coding tasks and general question-answering, the best open models now match paid services almost exactly. The remaining gap, roughly 3 percentage points on standard tests, concentrates in complex multi-step reasoning. For the overwhelming majority of business tasks, that gap does not matter.
The cost difference is another matter. On a neutral test bench, the strongest open model costs about one-fifth what the leading paid service charges for the same task. For a business running thousands of AI tasks a day, that ratio determines whether AI is a manageable line item or a budget crisis.
Microsoft found out what a budget crisis looks like. The company gave thousands of its engineers access to Anthropic's Claude Code in December 2025. The tool was genuinely popular. It was also paid per use, and by mid-2026 the token bills had consumed the annual AI budget ahead of schedule. Microsoft cancelled most licences by June 30 and shifted engineers to its own internal tool. Uber had a parallel experience: it deployed the same tool to 5,000 engineers, burned through its entire 2026 AI coding budget in four months, and capped individual spending at $1,500 per person per month to stop the bleeding.
Stripe went the other direction. The payments company switched to running open models on its own infrastructure and cut inference costs by 73%, serving the same volume of API calls on one-third the computing hardware.
Then came the event that made the strategic picture impossible to ignore. On June 12, 2026, the US government ordered Anthropic to shut off its two newest models for all foreign nationals, everywhere in the world, within hours. Anthropic had no real-time way to verify nationalities, so it pulled the models for all customers globally. Every business that had integrated those models into live processes inherited a shutdown they played no part in and had no warning of. The controls were lifted about two and a half weeks later, but the lesson was clear and permanent: a rented service comes with an off switch that belongs to someone else.
Open-source models cannot be switched off this way. Once the model is downloaded and running on hardware you control, no government order, vendor pricing change, or acquisition reshuffles your operation. That is the value of ownership, and it is not an abstract point anymore.
The challenge with open models is not quality. It is the operational work of running them. A paid API takes an hour to connect; a self-hosted model takes weeks to deploy properly, and ongoing maintenance falls on your own team. That gap is real, and it is why large enterprises still lean on paid services even when they are more expensive. But the gap is closing as the tools for running open models become more reliable.
The geography of this shift is also worth noting. Chinese open-source labs, particularly Alibaba and DeepSeek, now produce the most-downloaded models in the world. Alibaba's Qwen family has been downloaded 942 million times from the main model-sharing platform. These models are not restricted to China. They run on hardware anywhere in the world, and restricting the hosted web service, which several governments have done, does not stop anyone from downloading and running the weights locally.
For a business operator, the practical question is simple: which of your AI workloads are high-volume and repeatable enough to justify switching from a pay-per-use service to aself