When I started building Process Runner, the models it ran on were the best available. Most of them have since been replaced, some more than once. Prices changed. Limits changed. New providers showed up, and a few of the old leaders fell behind.
If Process Runner had been welded to any one of those models, every one of those changes would have been a rebuild.
In AI, the model is the part most likely to change. Build so that changing it is a setting, not a rewrite.
The Landscape Moves Faster Than Your Build
Most software can count on its foundations holding still for years. A database or a programming language changes slowly and warns you first.
AI models don't work that way. A new release can make last quarter's best choice the expensive, slower option. A provider can retire a model you depend on, change its pricing, or hit a rate limit in the middle of your workday. None of that is a failure on anyone's part. It is simply the pace of this field right now.
Even the race between the big labs' closed models and openly released ones keeps shifting. Stanford's 2026 AI Index found the best open model within 0.5% of the best closed model in August 2024; by March 2026 the gap had widened again to 3.3%. Epoch AI puts the best open models about four months behind the frontier. Whichever side leads next year, a model-agnostic build can use it.
A build that assumes one model will stay on top is a bet against that pace.
People Are Already Switching
This is not theory. OpenRouter is a marketplace where developers and apps reach hundreds of models through one connection, so its traffic shows what people actually choose. From late 2024 to late 2025, closed models from the big labs carried about 70% of its token volume, and openly released models about 30%. By June 2026 that had flipped: 69.1% of tokens went to open-weight models and 30.9% to closed ones.
The money moves more slowly. Through much of 2025, closed models still took roughly 96% of the revenue, because the hardest work is still worth paying a premium for. What people are changing is where the everyday, high-volume work goes.
That move takes a build that can make it. If every call in your system is written for one vendor, cheaper open models are not an option for you. They are a rewrite.
The Process Is the Asset. The Model Is a Worker.
The most important decision in Process Runner was where the value lives. It does not live in the model. It lives in the process.
A process is a file: the steps, the rules, the checks each step must pass, and the evidence it has to produce. When a step needs AI, it asks for what the work requires: strong reasoning, low cost, speed, a long document read. It does not ask for a brand name. The router picks the model that fits.
The knowledge works the same way. What a step needs to know is written down in files any model can read, not buried in one vendor's prompt format or tools. When a better model arrives, it inherits all of that on day one.
What Being Agnostic Bought Us
Resilience. At one point, a rate limit at one provider stopped a multi-step run halfway through. The lesson was not "pick a better vendor." It was that no single provider should be able to stop the work, so a step has to be able to move to another model and keep going.
Cost. Not every step needs the most powerful model. In one test, a free model running on a local machine drafted a solution and a frontier model fixed only what was left. The result passed every check for less than a penny. You can only route by difficulty if you are free to use more than one model.
Quality. Different models make different mistakes. Having a second model, from another vendor, review the first one's work catches problems the author would not see in its own output. That only works if the system can talk to more than one vendor.
Build for the Model You Haven't Met Yet
The best model today is a snapshot, not a foundation. The right one for your work next year may not exist yet.
So build for it now. Keep the process, the rules and the knowledge yours, and treat each model as a capable worker you can hire for the job in front of it.
Don't marry the model.
Hire it.
Sources
- OpenRouter, "State of AI: An Empirical 100 Trillion Token Study" (December 2025)
- Tomasz Tunguz, "The Thriving Ecosystem of Open Models" (June 2, 2026)
- Tech Times, "Chinese Open-Weight AI Handles Most Developer Tokens: Closed Models Capture 96% of Revenue" (September 17, 2026)
- Stanford HAI, "The 2026 AI Index Report: Technical Performance"
- Epoch AI, "Open models lag state-of-the-art closed models by 4 months" (May 29, 2026)