On June 2, 2026, at Build, Microsoft shipped seven of its own MAI models, trained in-house specifically to lean less on OpenAI and cut developer costs. One of them, MAI-Code-1-Flash, is a 5-billion-parameter coding model that Microsoft says solves harder coding tasks with up to 60% fewer tokens, and it is already rolling into GitHub Copilot and VS Code. Another, MAI-Thinking-1, is a 35B reasoning model trained from the ground up with no distillation from OpenAI's models [1][2].
The headlines are about Microsoft versus OpenAI. The thing that actually matters for anyone who builds on these models is quieter, and it is this: the model under your application is now a swappable commodity, and the rate at which it swaps is accelerating. The durable asset is not the model. It is your prompts.
I want to make that concrete, because it sounds like a slogan and it is actually an operational claim with consequences for how you work this month.
Count the model releases since mid-April
If you maintain anything on top of an LLM, your last six weeks looked like this: Claude Opus 4.8 (May 28), Google Gemini 3.5 Flash GA (May 19) with Gemini 3.5 Pro due any day, OpenAI's GPT-5.5 family with GPT-5.6 already announced for June, DeepSeek V4, and now seven Microsoft MAI models in a single keynote. That is not a frontier moving once a year. That is a frontier that re-prices and re-ranks itself every couple of weeks.
When releases came annually, "which model" was a strategic decision you made once. When they come fortnightly, "which model" is a routing decision you make continuously, per task, per price point, per benchmark. MAI-Code-1-Flash exists precisely so Microsoft can route Copilot traffic onto a cheaper model when it is good enough. You will do the same thing with your own workloads, whether you plan to or not, because the economics force it.
Commodity is the right word, and it is not an insult to the models. It means they are increasingly interchangeable at the interface, differentiated mostly on price, speed, and a few benchmark points, and replaced often. That is a healthy market. It is also a problem for the layer that sits on top.
What commoditization does to a prompt
Here is the part people underrate. A prompt is not portable for free. It is tuned, explicitly or implicitly, to one model's disposition, and that tuning does not transfer cleanly when you swap the backend.
I have written this same warning three times in the last two months, because the same thing keeps happening with every release:
- Opus 4.8 started flagging uncertainty and refusing to rubber-stamp, so prompts written to extract one confident answer suddenly behaved differently.
- Gemini 3.5 Flash silently moved its default reasoning level, so prompts that did not set it explicitly ran weaker than the week before.
- GPT-5.5 calibrated verbosity and instruction-following differently, so step-by-step scaffolding that helped the old default started to hurt.
Now multiply that by a world where you might run MAI-Code-1-Flash for cheap bulk coding, Opus for the hard reasoning, and Gemini for long-context, all in the same product. Every one of those backends wants its prompt shaped slightly differently. A prompt that is excellent on one is merely adequate on another and occasionally broken on a third. Swapping the model is cheap. Re-validating every prompt against the new model is the actual cost, and it is the cost nobody budgets for.
Concretely: a JSON-extraction prompt that was rock solid on one model can start wrapping its output in a sentence of prose on a model that calibrates verbosity differently, and your parser breaks in production before anyone clocks that the model even changed. The swap was a one-line config edit. The fallout was a downstream incident. That asymmetry, trivial to swap and expensive to verify, is the entire problem in one line.
So the question stops being "which model is best" and becomes "how do I move my prompts across models without silently degrading." That is a prompt-management question, not a model question.
The moat is the prompt layer, and it has to be deliberate
If the model is the commodity, your defensible asset is the layer above it: the prompts, the versions, the knowledge of which prompt works with which model and why. That asset only exists if you build it deliberately. Most teams do not. They keep prompts in a dozen places, paste them between tools, tweak them per model, and lose track of which version was tuned for what. When the next model drops, they re-discover the same fixes from scratch.
Here is the model-agnostic discipline I now run, and it is the difference between a model swap being an afternoon and being a regression hunt:
-
Keep prompts model-independent by default. Write the intent and the constraints, not the model-specific scaffolding. The "think step by step" and "always respond in JSON, no preamble" hacks are compensation for a specific model's quirks; keep them in a clearly marked layer you can strip or swap, not baked into the core prompt.
-
Version every prompt with its target model noted. When you tune a prompt for Opus 4.8 or MAI-Code-1-Flash, record that. The version is worthless six weeks later if you cannot tell which model it was shaped for.
-
Score and regression-test on every swap. Before you route a workload to a cheaper model because it is "good enough," measure it on your own representative inputs, not on a benchmark. Good enough on SWE-bench is not good enough on your codebase.
-
Keep one source of truth, not a dozen pasted forks. The moment the same prompt lives in three repos and two chat tools, you have lost the ability to update it once and trust it everywhere. That is also a security surface, as anyone who read the agentic prompt-injection piece now knows.
-
Treat the model as a backend you will replace. Architect so that swapping MAI-Code-1-Flash in for a pricier model is a config change plus a test run, not a rewrite. If swapping the model means rewriting prompts everywhere, your prompts were too coupled to begin with.
None of this is exotic. It is the same discipline software teams learned about dependencies a decade ago: pin what you depend on, abstract the swappable parts, test before you upgrade. The model is now a dependency that ships a new major version every two weeks.
Why this release in particular makes the point
Microsoft is the most integrated AI vendor there is, and even Microsoft decided it did not want to be locked to a single model provider. It built its own to gain optionality and lower its bill [1]. That is the tell. If the company with the deepest OpenAI partnership in the industry is engineering itself the freedom to swap models, the freedom to swap models is the strategic position, not loyalty to any one model.
You do not have Microsoft's budget, but you can have the same posture at your scale, and the cost of it is mostly discipline. Keep your prompts portable, versioned, and tested, and every new model, MAI, GPT-5.6, Gemini 3.5 Pro, becomes an opportunity to cut cost or gain quality rather than a fire drill. Skip that discipline and each release is a tax.
This is the problem Keep My Prompts is built around. Every prompt is versioned with a note on the target model, the Prompt Score evaluates it on six criteria so you can compare the same prompt across models before you switch, and you keep one authoritative copy instead of forks scattered across tools. When MAI-Code-1-Flash or GPT-5.6 lands and you want to know whether to route to it, you test your actual prompts against it and see the score move, instead of guessing. Free to start, no credit card.
It pairs with the broader case for choosing models as a continuous decision rather than a one-time bet: the multi-model world is here, and it rewards the teams who treat their prompt library as the stable layer.
The signal
Seven models in one keynote is not the story. The story is that "seven models in one keynote" is now a normal week, and it will keep being one. The model layer is consolidating into a commodity that re-prices constantly, and the value is migrating up the stack to whoever holds a clean, portable, tested set of prompts.
Pick your model for today's task on today's price and benchmark. But build as if you will replace it next month, because you will. Version your prompts, note the target model, test before you swap, and keep one source of truth. The model is rented. The prompt library is owned. Act accordingly.
Keep My Prompts helps you version prompts per model, score them on 6 quality criteria, and move them across models without silent regressions. Free to start, no credit card required.
References
[1] Microsoft unveils new AI models to lessen reliance on OpenAI and lower costs for developers, CNBC, June 2, 2026. https://www.cnbc.com/2026/06/02/microsoft-unveils-new-ai-models-lessen-reliance-on-openai-lower-costs.html
[2] Biggest Microsoft Build 2026 announcements: MAI models (MAI-Code-1-Flash 5B, MAI-Thinking-1 35B), Tom's Guide, June 2, 2026. https://www.tomsguide.com/news/live/microsoft-build-2026
