Vendor lock-in in LLMs rarely comes from pricing. It comes from integration surface area — the accumulated code that assumes one provider's request shapes, error codes and streaming semantics.
How lock-in actually accumulates
It starts small. A retry wrapper that checks a provider-specific error code. A token counter using that provider's tokenizer. A streaming parser that expects their SSE event names. None of these are decisions anyone made deliberately; they're the natural residue of shipping.
Six months later, switching providers means auditing every one of those touchpoints, and nobody wants to fund that project.
The cheap insurance
Keep the interface standard. If your client code targets an [OI]-compatible surface, the provider becomes a configuration value rather than an architectural commitment. Model ids are the only thing that changes.
Isolate provider specifics. Anything that must know about a specific vendor — tokenizers, special parameters, error mapping — belongs behind one adapter, not scattered across features.
Rehearse the switch. Once a quarter, route a non-critical endpoint through a second provider. If it takes an afternoon, you're not locked in. If it takes a sprint, you've found your real risk.
Pricing transparency as a signal
A gateway that publishes full per-token rates is easier to leave, which sounds bad for the gateway and is actually good for you. Our rate card is public: KeyoAPI [OI] rates.
The point isn't that you should switch often. It's that the ability to switch cheaply is what keeps pricing honest — and makes the one time you do need to move an afternoon instead of a quarter.
Top comments (0)