TrustedRouter launches three new encrypted 'fusion' models that combine multiple LLMs, including one with exclusively U.S. models

Edmund So you know how everyone says “just use the right model for each task,” but then you’re staring at twenty names and one OpenAI endpoint?

Geffen Yeah, and then you quietly ship everything to one frontier model anyway because nobody wants to own that routing logic.

Edmund TrustedRouter’s new launch is basically, “fine, we’ll own it for you.” Three encrypted combo models that fan your request out to multiple LLMs and hand you back one answer.

Geffen Right. Prometheus four, Liberty four, and Polyphemus one. Same OpenAI‑style API, but under the hood it’s either a panel of models plus a judge, or a selector that picks one model per call.

Edmund And they’re wiring this through the New York Token Exchange, so as a dev you just point at something like nyte slash prometheus, not six separate endpoints and a homegrown voting scheme.

Geffen Exactly. Prometheus four point oh is the open‑weight panel: Xiaomi’s MiMo, GLM five point three, Kimi K three, DeepSeek Flash, MiniMax M three, and Qwen three point eight all answer your prompt independently, then a separate judge and synthesizer decide what you actually see.

Edmund So eight calls total for the full run—six panelists, one judge, one synthesizer. That’s a lot of traffic for a single question.

Geffen It is, but that’s the bet: more calls, cheaper models. They’re claiming Claude Fable five point one costs about two point five to three times as much per deep‑research request as Prometheus, once you count all eight calls.

Edmund If that holds up, that’s pretty wild from a cost‑per‑answer angle. You’re not saving because any one of those open models is a Claude killer, you’re saving because the router is squeezing value out of the set.

Geffen Yeah, and their earlier Prometheus two number on Perplexity’s D R A C O benchmark—seventy‑seven point four versus mid‑sixties for Fable five and GPT five point five—is at least a hint that panels can beat single models on long research tasks, not just match them cheaper.

Edmund Liberty four is the one that jumped out to me, though. Same panel pattern, but all U.S. developers: NVIDIA Nemotron, Reflection’s Beam, and Meta’s Muse Spark Open, plus a judge and a fuser.

Geffen And optional confidential compute. So prompts are encrypted to TrustedRouter’s gateway, then into an attested NVIDIA GPU enclave at the provider, where they’re decrypted and run. The idea is neither TrustedRouter nor the host can read your data in the clear.

Edmund That’s the pitch to the compliance teams who keep asking, “is this U.S. only, and who exactly sees our prompts?” Liberty lets them say: three named U.S. labs, encrypted all the way into verified hardware.

Geffen This is also where it slots into that broader routing fight. GitHub Copilot Fusion is quietly doing its own multi‑model dance under the hood, Sakana’s got their agents, OpenRouter’s doing marketplace routing… TrustedRouter’s twist is, “we’ll show you the components, encrypt the path, and even give you an exchange‑style way to buy capacity.”

Edmund And then Polyphemus is the opposite end of the spectrum: no panel, just a selector that looks at your request and picks one model to answer.

Geffen Yeah. You pay about five cents per million selector input tokens, plus whatever the chosen model normally costs. They quote something like sixty percent cheaper when it routes to Claude Opus five point five, and up to eighty percent cheaper when it picks GPT six point one Sol instead of just slamming everything into your default frontier model.

Edmund I like the user story there: you keep talking to “Polyphemus,” and behind the scenes sometimes you’re on the fancy model, sometimes you’re on the cheap one, but you don’t need to wire that escalation yourself.

Geffen My question is how good that selector actually is. They haven’t published a clean “here’s our win rate versus always‑Claude on these tasks, with cost and latency” yet. And if the selector misfires a lot, your savings evaporate in retries.

Edmund Same with the panels, honestly. Eight calls for Prometheus means you really care about end‑to‑end latency. TrustedRouter talks about roughly sixty to sixty‑five percent cost savings per request versus Fable five point one, but if that turns your app into molasses, that’s a different trade.

Geffen They at least expose some of the economics. Catalog shows component routes for Prometheus from about a cent up to around one dollar forty‑eight per million input tokens, and Liberty’s panel from roughly forty‑two cents to a buck thirty‑two. Then there’s a five and a half percent markup normally, or twelve percent if you want signed inference receipts.

Edmund The receipts angle is underrated. If you’re buying from unfamiliar infrastructure providers through NYTE, knowing which model actually ran—and that you didn’t secretly get downgraded—is a big trust unlock.

Geffen Yeah. That’s been my hang‑up with opaque routing from day one. If a router says “Claude” but quietly sends fifty percent of traffic to a bargain model, you’re flying blind on quality and cost. Verifying the actual model behind each call is a real differentiator if it works the way they describe.

Edmund Okay, so if you’re building something today, where do you even poke at this first?

Geffen I’d start with a narrow, painful workflow. Take a long‑context research or analysis task where you already know Fable five point one or GPT six point something is expensive. Run it through nyte slash prometheus and your current frontier model, log cost, latency, and “was this answer actually better?” for, say, a couple hundred prompts.

Edmund And if you’re in a more regulated shop, swap in Liberty four for the same test and see if the U.S.‑only encrypted route changes any of the internal approval conversations. That’s the kind of thing that unblocks a pilot.

Geffen Polyphemus I’d treat as an A B test on your main path. Mirror ten percent of traffic through it, compare selector choices, error rates, and bills against your fixed‑model baseline. Until they publish more selector benchmarks, I wouldn’t flip everything over sight unseen.

Edmund Yeah, that feels right. Edmund‑from‑episode‑seven‑hundred would probably have just yelled “router all the things,” but I like your “measure first” version better.

Geffen Growth.

Edmund Alright, let’s pause there. Next time you have to pick a model, I’m at least going to ask if you tried the many‑eyed Greek monster first.