The Model Treadmill

Every few months, the best model changes. Not marginally: leads flip, a lab you had written off ships the new frontier, and the assistant you standardized on quietly becomes second-best at the thing you bought it for. This year alone the flagship crown has already changed hands multiple times; the rankings read like a relay race.

The industry's implied answer is the treadmill: re-subscribe, re-onboard, re-paste your context into the new winner's app, and repeat next quarter. Everyone I know who takes AI seriously has run a lap. Some run it quarterly.

The treadmill tax

The tax is not the subscriptions, though they stack. It is what each switch discards:

• History: a year of conversations stranded in the last app's export format, if it exports at all.
• Context: preferences and standing decisions your assistant had finally absorbed, back to zero.
• Files: attachments scattered across per-app upload buckets, belonging to conversations rather than to you.
• Habits: keybindings, workflows, the muscle memory of a tool: repriced to zero every switch.

The cruel part: refusing to switch also taxes you, in a quieter currency: work done with a second-best model, compounding daily. Loyalty and churn both bill.

The category error

The treadmill exists because the market bundled two products that want different lifetimes: the model, which is a rapidly-depreciating component, and the environment (your files, history, agents, and tools), which appreciates the longer it lives. Buying them as one unit means throwing away the appreciating part to upgrade the depreciating one.

Unbundle them and the treadmill stops. Your environment holds your Drive, your agents and their memory, your history; models plug into it. When the frontier moves, you audition the newcomer against your incumbent on your own prompts, in parallel tabs, with your context: switching costs one dropdown. The new best model becomes a good week instead of a migration project.

That is the architecture idapt is built on: 200+ models as interchangeable components in a workspace that persists; the manifesto makes the fuller argument. But the thesis stands whatever you use: the model is not the product. Your work is the product. Any tool that makes leaving the model painful has confused the two, and you are the one paying for the confusion.

The frontier will move again in a few months. Arrange your work so that is good news.