Teams that never tested a second AI model are exposed if access to the first is cut

Pamela Fox, a Principal Cloud Advocate at Microsoft, argued on the AI Engineer show that sticking with one AI model is a choice with a price. Teams that never ran their app on another model would struggle if access vanished. She also showed a method that tunes the prompt to a written spec, so a cheaper model can do the job.

What was said

In a workshop, Fox ran identical tasks on four models she had set up in her Foundry project at Microsoft: Mistral Large 3, DeepSeek V4 Flash, Kimi K2.6 and GPT-5.5. Her reason was blunt. She warned that the US government could shut off access to a model at any point. Her conclusion was that "we need to have options."

The first test asked each model to count the letter E in one sentence. GPT-5.5 and Kimi both answered 13, which she called correct. DeepSeek V4 Flash said six. Mistral said eight. Fox said benchmarks only give a rough idea, and that you learn how a model performs by using it on your own task.

Later she showed DSPy, a tool that writes the prompt for you. You supply checks the output must pass, and an optimiser searches for a prompt that meets them. It took about six minutes on her task. She tried it on Mistral because it struggled most. She said it "can give you a lot more model independence." She also said Dropbox and Shopify had given talks on using this approach to move to smaller models.

Why it matters

Fox's approach changes what the team owns. The written checks are the lasting asset, and the prompt can be regenerated for whichever model is cheapest. Model choice then looks more like a price decision than a foundation. For buyers, a simple question follows: has anyone run our app on a second model, with checks that say what good looks like?

The other side

Fox presented this as an option, not a cure. She called it "potentially" powerful, and she showed one task. In the excerpts she gave no cost figures and no final scores for the optimised prompt, only that it worked with Mistral.

Swapping is also not free. She noted that Anthropic's models use a different API, so code has to change. The method also needs evaluations to exist, and she admitted they are hard to set up before launch. She said people are told to run them and few do.

Her cut-off warning was a general caution, not a forecast. She offered no odds, and the conversation did not say which models or customers would be affected.