Host: Whenever people ask what is actually powering myOrbit under the hood, they usually expect one of two things: either we are running some cheap open-weight model in a basement, or we are locked into a single tech giant.

Listener: Right, because usually an AI app just picks OpenAI or Claude and sticks with it. So what is the straight answer?

Host: The straight answer is four frontier labs: Anthropic, OpenAI, Google, and xAI. Anthropic, OpenAI, and Google power the core systems, Aura and Echo OS, while xAI powers some of our avatars and is currently being evaluated for the core OS as well.

Listener: Wait, all four? Why juggle multiple labs instead of just picking the one that wins the benchmarks?

Host: Because no single model wins across the board, and the leader moves almost every couple of weeks. If you build an entire workflow on just one provider, you are stuck with their blind spots and their downtime.

Listener: That makes sense, but isn't routing between models mostly about shaving down your cloud bill? Like, send the easy stuff to the cheapest model to save money?

Host: That is how most routers work, like OpenRouter or FrugalGPT. For a generic developer tool, prioritizing price makes total sense. But our routing engine, Layer Cake, flips the priority order.

Listener: Flips it how?

Host: The first question isn't what is the cheapest model that won't totally fail. It asks: what does this specific person need from this specific task, and which lab is best at delivering it? Cost matters, but it comes second.

Listener: Give me an example of how that actually plays out in practice.

Host: Take our Researcher feature. One sub-question might go to Anthropic for structured nuance, another piece goes to Google, and then the final claims are fact-checked by a completely different lab than the one that drafted them.

Listener: Oh, cross-checking across different companies so they catch each other's mistakes. That's clever.

Host: Exactly. And in everyday chat, inside Aura, Layer Cake sizes the thinking required for each turn. Most people pick a heavy reasoning model in an app and leave it on, so the model spends thirty seconds of expensive deep thought answering what time their dentist appointment is.

Listener: Guilty as charged. I do that all the time.

Host: Everyone does! Aura automatically takes a fast path for quick lookups, a normal one for daily work, and brings in the heavy frontier reasoning only when the problem genuinely calls for it. You get speed and quality without having to flip toggles.

Listener: What happens when an API goes down? We all know these AI labs have outages.

Host: Every fallback chain in Aura crosses providers. If OpenAI has a bad afternoon, the request moves to Anthropic or Google instantly. It only stops if all three major labs go down at the exact same moment.

Listener: Does it ever mask code bugs, though? Like, if you send broken data, won't it just keep failing across all of them?

Host: We have a strict rule for that: failover only triggers on provider-side downtime or rate limits. If it is a malformed request on our end, it halts immediately. Better to fail loudly once than hide an engineering bug behind green dashboards.

Listener: And the rest of the stack? You mentioned search and voice earlier.

Host: Dedicated tools for dedicated jobs: Perplexity, SerpAPI, and Exa for research search, Replicate for media, and ElevenLabs for voice generation. It is all documented line-by-line in our public privacy policy.

Listener: So the big takeaway is you never have to guess which model to select or worry about who is leading the AI race this month.

Host: Precisely. The labs will keep leapfrogging each other indefinitely. Layer Cake is built so that when they do, your tools just get better automatically, without you having to rebuild your setup.