TrustedRouter bets on encrypted AI panels to cut inference costs

The AI market has largely trained enterprises to choose a primary model provider and build around it. TrustedRouter is proposing a different operating model: let an intermediary decide which systems should handle each request, and in some cases ask several models to solve the same problem before producing a final answer.

That approach could make model access more like cloud infrastructure, where routing, security, provenance and price matter as much as the performance of any individual model. But it also creates a difficult accounting problem. A cheaper model call is not necessarily a cheaper request if the system needs to make eight calls to produce one answer.

TrustedRouter said in its launch announcement that it has introduced three encrypted “fusion” models: Prometheus 4.0, Liberty 4.0 and Polyphemus 1.0. The products are available through the company’s API and through the New York Token Exchange, using model identifiers that begin with “nyte/.”

Prometheus takes the most expansive approach. It sends a request to six open-weight models, compares their independent answers, then uses a judge and a synthesizer to produce the final response. A complete run can therefore require eight model calls. The design resembles a review panel: several systems generate opinions, another evaluates them, and a final system turns that process into one answer.

TrustedRouter says Liberty 4.0 uses models from NVIDIA, Reflection AI and Meta. The company presents it as a U.S. provider only option under its routing rules. Its catalog identifies the component systems and says the route components resolve to providers recorded as U.S. based, according to TrustedRouter’s U.S. model catalog. Liberty generally uses five calls, making it less expansive than Prometheus while retaining the basic panel concept.

Polyphemus 1.0 is more conventional. Rather than asking multiple models to answer, it selects one model before generation. Its product documentation describes selector token pricing, API behavior and routing details, positioning it as a way to choose an efficient specialist without paying for a full panel. The Polyphemus product documentation is also important because it makes clear that routing itself is part of the product, not merely an invisible implementation detail.

The commercial argument is substantial if the estimates hold. TrustedRouter says its panel configurations can reduce request costs by about 60% to 65% compared with Claude Fable 5.1. It estimates that Polyphemus can cut costs by 60% to 80% in selected routing examples. Those comparisons are company estimates, not a complete matched evaluation across price, quality, latency and task success.

That qualification is central to the launch. A panel may improve reliability on research questions by exposing disagreements among models. It may also produce an apparently confident consensus when several component models share the same blind spot. More calls can increase the opportunity for verification, but they can also multiply failure points and make response times harder to predict.

The critical case is not simply that fusion is expensive. It is that the claimed savings and quality improvements are difficult to evaluate together. Sandrise’s review questioned whether Polyphemus has demonstrated a reliable selector win rate and whether Prometheus can deliver acceptable latency when a full run requires eight calls. Edmund and Geffen also argued that the published cost claims do not yet provide a clean comparison that includes quality and latency.

That criticism goes to the heart of enterprise adoption. A procurement team does not buy the lowest token price. It buys a system that completes a task accurately, quickly and predictably. If a routed answer saves 65% on inference but takes several times longer, requires additional monitoring and occasionally needs human review, the apparent savings may disappear in the total cost of operating the workflow.

There is evidence supporting the underlying thesis. The public TrustedRouter-Fusion-Draco repository presents reproduction and benchmark work arguing that multi-model fusion can outperform individual frontier models on deep research tasks. That supports the idea that independent answers can be useful, particularly when a question benefits from breadth and cross-checking.

But benchmark support is not the same as proof of a general business advantage. Results may vary by task, panel composition and judge quality. A synthesizer that is good at combining answers may still inherit errors from its inputs. A selector that performs well on one class of prompts may choose poorly when the request is ambiguous or outside the patterns represented in its evaluation data.

Privacy is the other major part of TrustedRouter’s strategy. The company says eligible confidential routes keep prompts encrypted until they reach verified hardware at the provider. It also says confidential computing and zero data retention can apply to component calls, judges, synthesizers and fallback systems.

TrustedRouter’s trust documentation describes its verified gateway and confidential-computing deployment, including a measured workload and mechanisms for customers to verify the running service. Signed receipts can identify the model and provider involved. That provenance could matter to regulated businesses that need evidence about where sensitive data went, rather than relying on a routing company’s general assurance.

Encryption does not remove every risk. Enterprises still need to understand which routes qualify for the protections, how fallbacks are handled and what metadata is exposed during orchestration. The more providers and components a request touches, the more important those operational details become.

AI News Radar’s neutral assessment says TrustedRouter’s quality, security and savings claims have not been independently verified and that matched evidence for cost, latency and task success remains unavailable. That is less a rejection of fusion than a description of the proof still required.

TrustedRouter’s launch therefore points to a broader shift rather than a settled product category. If routing systems can reliably select or combine models while preserving privacy, the winning layer in AI may sit above the model providers. It would manage the decision about which system should answer, prove what handled the request and optimize the total cost.

The unresolved issue is whether that layer creates enough value to justify its complexity. Enterprises will not adopt multi-model inference because eight calls sound more sophisticated than one. They will adopt it if the resulting answer is measurably better, the response remains fast enough and the savings survive the full cost of operating the system.