Ortelan launches Praus, a sovereign AI harness that routes each question to the right open model — and checks the answer

Ortelan Labs, a Salt Lake City applied AI lab, launched Praus on November 3, 2026, a sovereign AI harness for open-source models. Praus is not itself a model. It is the layer that sits between people and the models they use: for each request it chooses the models best suited to the task and checks the result before returning it. It handles research, writing and code, and its agents, now in preview for selected accounts, carry out multistep tasks.

The pitch is deliberately modest. Chief executive Kyle Nel put it this way: “Most people don’t need the biggest model for every question. They need the right one, checked, at a cost that makes sense.” That framing — routing plus checking, rather than one giant model for everything — is the whole product. Whether the checking is good enough to trust is the question the launch leaves open, because the evidence Ortelan published is its own.

The numbers, with the caveat attached

Alongside the launch, Ortelan published Praus’s results on nine public AI benchmarks on Hugging Face. In the company’s own testing, Praus ranked first among the listed systems on two of them, GraphWalks and SimpleQA, and between third and sixteenth on the other seven. The company also estimates its on-premises infrastructure cost for a typical answer at $0.00069, below the list-price cost of all 17 frontier models it compared. That is an infrastructure-cost estimate for a typical answer, not a price a customer pays — the release names no subscription price.

Read those figures as Ortelan’s account of its own system, not as an independent ranking. The release says so itself: the benchmark scores and cost figures are self-reported estimates from Ortelan’s own testing, have not been independently verified, and are not guarantees of performance, accuracy or price. The headline framing of “frontier-level” results is the company’s, and the release does not name the individual open-source models Praus routes between. Until someone outside Ortelan reproduces the results, the honest summary is that Praus looks competitive on two public benchmarks in its maker’s testing, and unproven everywhere else.

What Praus does day to day

Praus’s features cover the familiar assistant ground, with two ideas doing the extra work. Agents carry out multistep tasks using tools and keep a record of each step; coding assistance gives a reason for each change. Alongside those are web search and deep research reports, Projects that keep chats, files and notes together, group chats with friends, family or coworkers, and voice conversations with read-aloud answers. Developers can apply for access to the Praus API, which Ortelan opens account by account.

The sovereignty claim rests on where the models run and who sets the rules. Praus runs open-source models on Ortelan Labs’ own on-premises infrastructure, with additional capacity from cloud providers, and lets customers set the rules for how their data is used. Because it is not tied to any single model, Ortelan says newer open-source models can be added or swapped in as they improve, with no change for the people using it. Its agents work through tools that run in the customer’s own application or terminal, so every action takes place in the customer’s environment and leaves a result that can be checked.

Who is behind it, and what to watch

Ortelan Labs is led by chief executive Kyle Nel and partners Christopher Harper and Bradden Blair. Before founding Ortelan Labs, Nel was executive director of Lowe’s Innovation Labs. The lab also builds AI systems for clients and invests in early-stage AI startups, but Praus is the product it is putting in front of the public.

Praus is available now at praus.ai, and an iOS app is planned for later this month — a plan the release flags as forward-looking and subject to change, as is the preview status of the agents. What to watch is straightforward: independent benchmark results, the names of the models under the harness, and what Praus actually costs to use. On those three, the launch announcement is silent. The routing-and-checking idea is sound, and the self-reported numbers are encouraging, but a harness earns trust by being checked by other people. Praus has not been yet.