The OpenAI and Cursor cutoff exposed the gap between offering models and letting developers actually move between them. That gap only becomes visible when a model goes away, which is a bad time to find out where your tool put its seams.
We’ve made the case for independence and incentives already. This is the mechanical version: what a tool has to get right for switching to be viable during an event like this.
The cutoff wasn’t about anything developers did #
Cursor didn’t lose access for abusing the API or breaking usage limits. It lost access because SpaceX bought it, which triggered a change-of-control clause, and OpenAI pointed to X’s history with its contracts and to Musk’s sworn testimony about distillation as justification. Anthropic had already banned xAI earlier this year over similar terms violations, so this isn’t a one-lab quirk.
Whatever you think of the merits, the mechanism is what matters for planning. Your access to a model depends on a contract between two companies you don’t work for, and it can end for reasons that have nothing to do with your usage, your spend, or your behavior. There’s no version of being a good customer that protects you from it.
Pricing has to survive the swap #
Flat subscriptions quietly price in a specific model mix. When part of that mix disappears, you’re paying the same amount for a smaller menu, and the vendor has every reason to steer you toward whatever is cheapest for them to serve.
Kilo charges per token at the provider’s rate. Switching from Claude to Gemini to a Grok model changes what you pay because it changes what you used, and nothing about your bill depends on us preferring one lab.
Your context has to come with you #
Switching models is cheap. Rebuilding everything around them isn’t. If your rules, prompts, indexed repo, and session history are tied to a particular provider integration, then model freedom on paper is a migration project in practice.
Kilo Sessions persist across VS Code, JetBrains, the CLI, and the web, and they’re resumable and shareable from any of them. Codebase Indexing runs on cloud-hosted embeddings tied to your repositories rather than to a model vendor. You can change the model, and that layer doesn’t notice.
A vendor that makes models wants you using them #
The Hacker News read on this was blunt: labs will build strong harnesses for their own models while restricting access to competing ones, which leaves independent harnesses as the ones that keep working across providers.
You can see the same pressure inside the products. Memeburn reported that Cursor’s routing requires Grok 4.5 to function, and the HN thread from late August includes a complaint about Cursor changing a developer’s selected model to the newest Grok without asking. Neither requires bad intent. A company that trains models has a rational reason to make its own the path of least resistance, and that reason doesn’t go away when you’d rather use something else.
Kilo doesn’t train models. There’s no house model our routing protects, no lab whose roadmap sets our roadmap, and nothing we gain when you pick one provider over another.
How to prepare for a future event #
None of this makes a model permanent. Providers change terms, restrict access, deprecate, and get into fights with each other, and any lab can pull a model out of any tool the same way OpenAI pulled its models out of Cursor. What changes is the size of the hole it leaves. With 500+ models available and switchable mid-session, losing one is a rebinding, not a rebuild.
The teams that will handle the next case comfortably are the ones who can point their workflow somewhere else with two clicks.