One's how requests leave your cluster. The other's how your Next app talks to models. Different floors. Somehow this still starts fights.

I keep seeing the same argument online: LiteLLM or the Vercel AI SDK, pick a side, defend it like a football club.

That fight's mostly a category error dressed up as taste. Asher's framing on X (@Asher101M) matched what I've been muttering in reviews for months. These things sit on different layers. Shocking, I know.

LiteLLM (as I use the word here) is infra-shaped. You want one OpenAI-compatible HTTP front door. Lots of providers behind it. Keys, routing, fallbacks, spend controls: the glamorous work of making sure nobody commits a key to git and calls it "temporary".

The Vercel AI SDK is app-shaped. You're in TypeScript / Next. You want streams that don't feel cursed. UI hooks. Tool calling that lands in the product UX without reinventing half a protocol because someone watched a demo once.

If you're comparing them like two ORMs, you'll stay confused. One's closer to a reverse proxy for models. The other's closer to how your React tree consumes those models. They're not competing for the same gold star.

Building a Next product surface? I start in the AI SDK. That's where the craft is for frontend folk: tokens on the wire as a UX problem, not a YAML archaeology dig.

Need one gateway so staging and prod don't grow six SDKs and twelve env vars named OPENAI_KEY_FINAL_FINAL2? That's LiteLLM (or something in that job family) sitting behind the app.

I've also stacked them. App talks AI SDK. AI SDK talks a gateway. Gateway fans out to providers. Nobody's betrayed their tribe. You've just drawn a line between product code and provider chaos, which, historically, is how grown-ups avoid writing the same OAuth regret three times.

You don't have to use both. A tiny demo can live on the AI SDK alone. A thin internal service might only need the gateway. Not every repo needs a cathedral.

But the moment someone says "we picked LiteLLM instead of the AI SDK", I want a whiteboard and a calm voice. Instead of what, exactly? The streaming UI hooks? Or the multi-provider front door? Those aren't the same purchase, unless your architecture diagram was drawn in crayon.

Stop casting them in a cage match.

Put the gateway where infra belongs. Put the SDK where the app belongs. Your future self will waste less time on fake rivalries, and more time on the bugs that actually ship.