Stop Chasing New Models. Build Once and Access Them All
Mozilla.ai
A new gateway called Otari wants to fix the mess of hooking your app into every new AI model separately. Instead of rewriting code each time a better model drops, you point at one endpoint and swap models like a config setting.
Based on reporting by Mozilla.ai — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Every few weeks another LLM shows up promising to be faster, cheaper, or just smarter at reasoning, and every few weeks some engineering team quietly groans. Trying out that new model rarely means just swapping a name in a config file. It usually means a new SDK, new auth flow, new request format, new error handling, and a pile of provider quirks nobody asked for. Multiply that by five or six providers and you're not building an AI product anymore — you're maintaining a compatibility layer for other people's APIs.
This is basically dependency hell, version 2.0. Anyone who shipped software before Docker and lockfiles remembers what happens when tightly coupled systems evolve independently: one update breaks something unrelated, and "works on my machine" becomes a running joke. The industry solved that with stable interfaces — package managers, containers, consistent contracts. LLM providers, with their own SDKs, credentials, and billing quirks, are recreating that same fragmentation, just one layer up the stack.
Otari's pitch is to put a single OpenAI/Anthropic-compatible gateway between your app and everything else. One API key, one request format, and the gateway handles routing, credentials, and fallback behavior across more than 40 providers behind the scenes. Swapping models becomes a routing decision instead of a rewrite — the application layer stays untouched while the model layer shifts underneath it. That's a real decoupling of two things that shouldn't have been welded together in the first place: how your product behaves, and which model happens to be serving it this month.
There's a security and cost angle too, and it's not a minor one. Centralizing provider credentials behind workspace-scoped keys means your app never touches the actual provider secret, which shrinks the blast radius when (not if) a key leaks. And instead of reconciling five separate billing dashboards to figure out which team burned through the budget, usage and spend get tracked in one place, tied to actual requests rather than a mystery lump sum on an invoice.
None of this replaces judgment — someone still has to decide whether a lighter, cheaper model is good enough for a given workload, or whether reliability and latency numbers hold up. What changes is that the decision no longer requires an application rewrite every time. The model ecosystem keeps churning; the interface in front of it doesn't have to.
My take — AI-written commentary, not fact-checked reporting
This is the unglamorous infrastructure work that never gets a keynote slide but saves teams months of pain, and I'm glad someone's finally treating model access like the commodity plumbing it should be. The real risk in AI right now isn't picking the wrong model — it's building your entire product around one vendor's SDK and discovering a year later you're locked in by accident, not by choice. Gateways like this are how open-weight and smaller models actually get a fair shot against the frontier giants, since suddenly the tooling gap stops being a moat.
Read more about this at: Mozilla.ai
