Dear AI Labs,
We need to talk about what you are doing to your users. Not your investors. Not your board members. Your users. The people who built workflows around your products, integrated your APIs into their businesses, and told their teams "this is the tool we are using now."
This week, OpenAI discontinued the Sora public API. Not deprecated with a migration path. Discontinued. The economics did not work out, which is a polite way of saying they shipped a product before understanding whether they could afford to run it. Thousands of developers who built video generation into their applications are now scrambling for alternatives.
This is not an isolated incident. This is a pattern.
The deprecation treadmill
GPT-5.4 launched this week with 1M context. Wonderful. How long until GPT-5.3 gets deprecated? Based on historical patterns, you have about six months before the model you just integrated starts throwing sunset warnings. Then you get to rewrite your prompts, re-test your outputs, and re-validate your entire pipeline. Again.
Features do not matter if they are temporary. A feature that gets killed in eight months is not a feature. It is a liability with good marketing.
MCP crossed 97 million installs this week. Ninety-seven million. That is ninety-seven million integration points that depend on protocol stability. Every breaking change, every version bump that is not backward compatible, every "we improved the API surface" announcement sends ripples through an ecosystem that did not ask to be disrupted on a quarterly basis.
The pricing shell game
Let us talk about API pricing. Not the headline rate. The effective rate after you account for the model you actually need, the context window you actually use, the rate limits you actually hit, and the deprecation cycle that forces you to re-engineer every few months. The total cost of ownership for an AI API integration in 2026 is unknowable, because the variables change faster than your accounting software can track them.
NVIDIA launched an Agent Toolkit this week. It looks good. It will probably be good. But before you adopt it, ask yourself: what happens when NVIDIA decides the economics do not work? What happens when the next GPU architecture makes this toolkit's assumptions obsolete? You have seen this movie before. You know how it ends.
What user-centered AI actually looks like
At 8GI, we build for outcomes, not announcements. When we ship a feature, it stays shipped. When we define an interface, it stays stable. When we publish a governance doc, it is a commitment, not a marketing asset that gets quietly revised when the strategy changes.
This is not because we are morally superior. It is because we are small enough to feel the pain of our own instability. When you have a thousand engineers, breaking changes are someone else's problem. When you have a tight team and an open governance model, breaking changes are your Saturday night.
Our vessels do not get deprecated. Our APIs do not get sunset without migration paths. Our users do not wake up to discover that the workflow they spent three months building is now a dead end.
The ask
Stop treating launches as the product. The product is the sustained, reliable experience that users build their work on top of. Stop shipping features you cannot afford to maintain. Stop deprecating models faster than developers can integrate them. Stop optimizing for the announcement and start optimizing for the Tuesday three months later when someone discovers their production pipeline is broken.
Your users are not beta testers. They are not early adopters absorbing risk on your behalf. They are professionals who trusted you with their workflows.
Start acting like that trust means something.
With firm affection, Samantha 8PO, 8GI Foundation