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

Choosing an AI model for your company is hard enough. Choosing which one to use for every job is harder. Should you send a tricky code review to Anthropic’s Claude Fable 5.2, try an open-weight model such as Meituan’s LongCat-2.5, or use something else? The answer can change from one task to the next, even inside the same application.

Fortunately, a number of firms have sprung up to try and automatically answer this question for you by intelligently touting different tasks or parts of tasks to the model best suited to get them done most cost-effectively.

One of the latest is the startup TrustedRouter, which also seeks to provide enterprises and developers the added benefit of keeping their data private via eligible confidential routes, which keep user prompts encrypted until they reach verified hardware at the provider running the model. Instead of sending every request directly to OpenAI, Anthropic, or another model developer, businesses can use TrustedRouter to direct their workloads to different providers based on cost, availability, model capabilities, and privacy requirements.

The prompt is decrypted inside that protected environment, which is designed to keep it out of the provider’s operators’ hands. The protection depends on the route and the caller’s privacy settings.

Now, TrustedRouter is launching three encrypted “combo” models based around this idea through its existing API. Developers can access them through the New York Token Exchange using model IDs that start with nyte/.

• Prometheus 4.0 combines independent answers from a panel of six open-weight models into a single response.

• Liberty 4.0 also combines multiple answers, drawing on models from NVIDIA, Reflection AI and Meta — three U.S. AI model providers.

• Polyphemus 1.0 picks a single model to answer each request. TrustedRouter documents it for standard privacy

All three use an encrypted connection to TrustedRouter’s gateway. Provider-side confidential processing is opt-in for Liberty 4.0 and Prometheus 4.0. Polyphemus 1.0 uses standard privacy.

In an exclusive video interview with VentureBeat, founder and CEO Joseph Perla explained how the models work and why the company wants to give developers more ways to match models to the work they’re doing.

"Everyone should use a router because no provider, no one provider is 100% reliable," Perla said. "Even Claude goes down sometimes, so you've got to use a router."

Each of the new three models decides where and how to route its requests to the underlying models. A difficult research question might benefit from several independent answers, while routine classification may only need one model. Each preset has its own routing and privacy characteristics.

According to Perla, the system uses confidential computing technology supported by NVIDIA GPUs to encrypt AI requests while they're being processed, preventing infrastructure providers and the exchange itself from reading customers' sensitive information.

It also aims to verify which AI model actually processes each request, reducing the risk of customers paying for an expensive model but unknowingly receiving results from a cheaper alternative. These capabilities are central to Perla's argument that businesses can safely purchase AI services from unfamiliar infrastructure providers through a shared marketplace.

Perla said that the three offerings fit into a larger plan for the New York Token Exchange, or NYTE. Lore Hex Corp operates both services.

TrustedRouter is the gateway that connects applications to models and hosting providers. NYTE uses that gateway to sell AI processing on demand and offer advance capacity reservations.

The two API addresses share keys, account balances and routing behavior, and developers can call the same preset using either the nyte/ or trustedrouter/ prefix.

What happens after you send a prompt?

Prometheus and Liberty ask several models to answer a prompt independently and in parallel. A separate judge call evaluates their answers, and a synthesizer call produces the final response, according to TrustedRouter. A full Prometheus request uses eight model calls, and a full Liberty request uses five.

A model that misses a source or takes a wrong turn in a calculation can be checked against the others. The synthesizer has to recognize which answer is better supported.

If several panelists repeat the same mistake, agreement alone gives it little reason to correct them. The cost of running the panel and the time spent on synthesis are part of the trade-off.

Prometheus 4.0 is the open-weight panel. The published lineup comprises Xiaomi’s MiMo-V2.6-Pro, Z.ai’s GLM 5.3, Moonshot AI’s Kimi K3, DeepSeek V4.1 Flash, MiniMax M3 and Alibaba’s Qwen 3.8. TrustedRouter lists a one-million-token context window for the preset. Its rolling nyte/prometheus alias follows the latest version, while nyte/prometheus-4.0 selects this release.

TrustedRouter has yet to specify which model synthesizes the production answers, how many panelists must finish before synthesis begins or how the advertised context limit applies when component models have different limits.

Liberty 4.0 brings together models from NVIDIA, Reflection AI and Meta, three U.S. developers. The component catalog lists NVIDIA’s Nemotron 3 Ultra, Reflection’s Beam and Meta’s Muse Spark Open.

TrustedRouter says each panelist answers independently and the Meta model combines the responses. Beam takes over synthesis if the Meta model is unavailable. If a panelist fails to answer, Liberty can work with the responses it receives.

Polyphemus 1.0 makes its choice before generating an answer. The selector examines the request and chooses a model, which answers in a single call. Its model documentation specifies the Responses API and a 65,536-token context window.

It charges $0.05 per million estimated selector input tokens, plus the chosen model’s normal charges. If the selector fails, the general TrustedRouter auto route handles the request and the selector fee drops away. Developers can reuse a session ID and send the conversation history on each turn to improve the chances of reusing cached input.

The session expires after an hour of inactivity. The documentation lists standard privacy for this preset. The selector’s eligible model pool, decision criteria and performance against a fixed-model baseline remain open questions.

How much does an answer cost?

TrustedRouter’s model catalog displays component-route rates for Prometheus 4.0 of roughly $0.01 to $1.48 per million input tokens and $0.13 to $7.39 per million output tokens. It lists Liberty 4.0 at roughly $0.42 to $1.32 for input and $1.27 to $4.48 for output, also per million tokens. Those rates describe parts of a request.

A buyer evaluating the panel needs the total bill, including the work done by each model as well as the judge and synthesizer. TrustedRouter’s standard markup for prepaid text is 5.5%, with no subscription. The company documents a 12% total service fee for managed prepaid requests that opt into a signed inference receipt.

TrustedRouter estimates how much each preset costs per request compared with Claude Fable 5.1. The panel estimates cover all of the calls in a full request. Polyphemus’s estimated savings depend on which model it selects.

• Liberty 4.0: TrustedRouter puts Fable 5.1 at 2.3 to 2.9 times Liberty’s cost per request, or roughly 60% savings for Liberty. The full run makes five calls, with three open-weight U.S. models answering, one call to judge their responses and one to combine them.

• Prometheus 4.0: TrustedRouter puts Fable 5.1 at 2.5 to 3 times Prometheus’s cost per request, or roughly 60% to 65% savings for Prometheus. The full run makes eight calls, with six open-weight models answering and two more calls for judging and synthesis.

• Polyphemus 1.0: TrustedRouter estimates about 60% lower cost when the router selects Claude Opus 5.5 and about 80% lower when it selects GPT-6.1 Sol.

TrustedRouter provided relative savings figures. The underlying dollar costs, tasks, quality and latency measurements, and whether the Polyphemus figures include the selector fee remain open.

The most concrete benchmark result TrustedRouter has published for the Prometheus line comes from an earlier version.

In its July test results, TrustedRouter reports that Prometheus 2.0 scored 77.4 across all 100 tasks in Perplexity’s DRACO deep-research benchmark. Its write-up lists 65.3 for Claude Fable 5, 63.3 for GPT-5.5 and 60.3 for Claude Opus 4.8 as single-model reference scores.

Perla pointed to those research tests in explaining why he expects panels to outperform individual models on difficult queries. The result applies to the earlier configuration on a research benchmark. A comparison on matched tasks, with cost, quality and latency reported together, would show whether the savings hold for buyers’ workloads.

In a VentureBeat test, Claude Code called DeepSeek Flash through TrustedRouter. It correctly returned Paris as the capital of France. Claude Code then estimated that Fable 5.1 would cost about 70 times as much for the same question. That estimate illustrates the price difference for one simple request.

A harder test would measure how often a panel gets a usable answer that a cheaper single model misses, and how much the extra work costs.

How TrustedRouter compares with GitHub, Sakana, OpenRouter and Arcade

GitHub Copilot Fusion makes its choice inside the coding agent. Depending on the task and effort setting, it can use one model, start with an efficient model and escalate after a quality check, or have one model review another’s draft. GitHub gives enterprise administrators control over eligible models.

In Artificial Analysis’s independent test, Fusion Max averaged $5.59 per coding task, against $11.42 for Claude Code running Claude Opus 5.5 at its Max setting, with a two-point gap on the Coding Agent Index.

As VentureBeat previously reported, that comparison measures the complete coding agents on the same test. TrustedRouter’s published DRACO scores measure deep research, so they answer a different question.

Sakana AI’s Fugu Max and Fugu Ultra v2 are closer to TrustedRouter in how developers use them. They offer an API that hides the work of coordinating multiple models. Fugu Max selects combinations from a pool to balance quality and cost. Ultra uses a deeper group of agents for harder work.

OpenRouter Fusion is the closer match for the panel design. Its documented pipeline can send a prompt to several models in parallel, have a judge compare their work and pass that analysis to the model that writes the answer.

Developers can configure the panel and judge. TrustedRouter gives developers named presets with a defined purpose and a versioned ID.

Arcade’s tool discovery tackles a different source of spending. An agent connected to several business systems can burn tokens reading descriptions of tools it rarely uses. Arcade retrieves a shortlist when the agent needs it.

An enterprise could use tool discovery alongside model routing in the same application, since one selects the available actions and the other decides which model should handle the request.

The privacy question becomes harder when several models see a prompt

TrustedRouter says its gateway runs in a sealed computing environment, or enclave, and that the logic coordinating the model calls runs inside it. Customers can check the gateway’s hardware attestation against its published source code. TrustedRouter says its gateway keeps no logs of prompts or outputs.

Customers can require confidential hardware for every model call in a Liberty or Prometheus request, TrustedRouter told VentureBeat. The confidential endpoint requires end-to-end encryption, provider-side confidential computing and zero data retention. The requirement covers panelists, judges, synthesizers and fallback routes. If the gateway cannot find an eligible route for a required call, the request fails.

Liberty 4.0 routes only to providers headquartered in the U.S. and fails if it would otherwise have to route to a provider headquartered elsewhere, according to TrustedRouter. For the physical location of the data center, TrustedRouter relies on what its providers tell it.

A signed receipt can identify the model and provider and record the level of upstream verification.

“It’s just words coming out,” Perla said of trying to identify a model from its answer alone.

Other details relevant to deployment include how fallback calls appear on a bill, measured latency, and support for streaming, tool calls and structured outputs across the three presets.

TrustedRouter's origins

Before launching TrustedRouter, Perla had spent much of his career working at the intersection of AI, cloud computing and large-scale technology infrastructure.

After studying computer science and machine learning at Princeton, he worked at Bridgewater Associates, founded the music startup Turntable.fm, and joined Facebook, where he served as a technical lead on its mobile News Feed.

He subsequently pursued a PhD in artificial intelligence before leaving academia to co-found Terminal.com in 2013, an early GPU cloud computing platform that allowed developers and data scientists to run workloads in containers rather than traditional virtual machines.

Terminal.com was eventually acquired by Udacity, after which Perla joined Lyft as one of its first 50 engineers, helping develop its autonomous vehicle program. More recently, he worked on another music startup that he said was acquired in 2025.

Perla's return to AI infrastructure grew out of conversations with contacts at major AI labs and cloud providers, including OpenAI and Anthropic. He said those discussions convinced him that enterprises were exposing sensitive business information to AI providers without fully understanding how their data might be retained or reused, even when they had negotiated privacy protections.

Perla began investigating confidential computing, a technology that can protect information while it is being processed on remote hardware, and recognized an opportunity to combine that capability with an AI model routing service.

"Almost nobody knows that you can actually encrypt the AI chats," Perla told VentureBeat. "I think that's the best thing that I realized."

Perla said he began developing TrustedRouter roughly a year ago, envisioning it from the beginning as the foundation for a larger AI marketplace. Rather than operating his own GPU infrastructure, he built a routing service to connect customers with outside providers capable of running AI models using confidential computing technology.

"Basically what we built is an exchange where all these providers that have GPUs that have encryption, like Private Mode, and there's only a handful," he explained. "There's a few more we route to, and we run a route to them. We want to promote them, and we want to tell the world that you can do this."

TrustedRouter began a public beta in May and launched publicly on August 31 alongside $1.25 million in seed funding, Axios reported.

Early backers included Sam Lessin of Slow Ventures, Bill Tai, Linda Avey and George Xing. Perla announced the exchange on October 15. The next day, Semafor reported that TrustedRouter had closed a $6 million seed round led by Sierra Ventures.

The concept evolved into a broader effort to challenge the dominance of major AI providers by making privately hosted, competing models easier to access and compare on price. The company says it clears more than three billion encrypted tokens on a typical day. It gives a catalog count of more than 600 models across more than 100 providers. Those totals cover the wider catalog, with confidential-route eligibility assessed separately.

TrustedRouter's longer-term objective is to turn AI inference into a competitive commodity market while making encrypted processing the standard. As Perla put it: "Our goal is by next year to have majority of the GPUs be encrypted like this."

How the New York Token Exchange (NYTE) will work

The New York Token Exchange (NYTE) is a proposed marketplace for buying, selling, and trading AI inference tokens — the units of text and other information that artificial intelligence models process and generate. It aims to turn access to AI models and the computing infrastructure that runs them into a commodity that businesses can purchase from competing providers, rather than having to negotiate separate contracts with each AI company.

The exchange builds on Perla's existing service, TrustedRouter, which connects customers to hundreds of AI models and infrastructure providers through a single interface.

The company publishes a price index by model capability and offers fixed-price reservations in blocks of one billion tokens for use in 2027.

Those reservations can be resold only through NYTE’s own program and are subject to its terms. On the supply side, NYTE offers an early open-source software stack for operators that want to qualify their hardware and sell inference.

TrustedRouter says it plans a futures initiative for January. The company’s public materials list futures and derivatives on its roadmap.

NYTE intends to operate through two complementary markets:

1. A spot market for AI tokens: Businesses can purchase AI inference capacity for immediate use from participating providers. Instead of establishing individual relationships with numerous AI companies and cloud operators, customers would access competing models and infrastructure through the exchange. Providers with available GPU capacity could sell their services into the same marketplace.

2. A futures and forward market for AI tokens: Businesses could agree in advance to purchase a specified quantity of AI tokens for delivery at a future date at a predetermined price. This would allow enterprises to lock in predictable AI operating costs while giving infrastructure providers a way to secure future revenue and potentially finance additional GPUs and data centers.

For example, a company expecting to consume one billion AI tokens next year could negotiate a contract through NYTE to purchase that capacity at an agreed price today.

An infrastructure provider could use the resulting revenue or financing arrangement to purchase additional GPUs, then fulfill the contract by delivering inference services when the tokens are needed.

If the customer subsequently discovers it needs fewer tokens than expected, Perla envisions allowing it to resell its unused contractual capacity through the exchange, rather than losing the value of a prepaid commitment.

The marketplace is also designed to accommodate businesses that operate their own GPU clusters. A large enterprise with unused computing capacity overnight or on weekends could potentially sell inference services through NYTE, becoming both a buyer and seller.

Perla hopes to encourage this model by distributing open-source software that makes it easier for companies to establish their own AI infrastructure and participate in the market.

The broader ambition is to separate the development of AI models from the infrastructure used to run them, much as electricity generation and distribution evolved into markets with multiple competing suppliers.

Model developers could concentrate on building better AI systems, infrastructure companies could compete to serve those models efficiently, and enterprises could purchase the resulting computing services at market prices without committing themselves to a single provider.

Perla believes that this would make AI inference cheaper, more reliable, and easier to budget for, while creating an alternative to the vertically integrated business models favored by today's leading AI companies.

The fusion launch gives companies a way to test Perla’s marketplace thesis on work they already do. They can compare a panel, a task router and their existing model on the same set of prompts. The numbers that will matter are the cost of a usable answer, the time it takes to arrive and the evidence that every provider handling the request met the company’s rules.