A sovereign harness for open-source models, built for fast, cost-efficient answers shaped around your work.
The tools to find sources, examine evidence and develop your work.
Get a quick answer with sources, or a full report when the question is bigger.
Keep each project’s chats, files and notes together, ready when you come back.
Bring friends, family or coworkers into one conversation, with Praus there to help.
Fix a bug, learn a language or plan a build. Every change comes with a reason.
Give Praus a file or data task. It works in a private cloud workspace and checks saved results before returning them.
Listen to answers with text-to-speech, or have a voice conversation.
How Praus works
Each request is routed and checked with answer quality, cost and your organization’s data controls in view.
New open-source models can be added or swapped.
Each request goes to the models best suited to it.
Responses are checked before they are returned.
Permissions and data rules apply to every request.
We test Praus on public benchmarks with our production system and score each run with the benchmark’s official grader.
Full results on Hugging FaceGraphWalks
100.0%
Rank 1 of 16
Multi-hop traversal over graphs given as long edge lists.
Accuracy, percent. Higher is better.
Estimated cost per general answer
$0.00069
Lowest of the models compared
Praus’s running cost is lower than every model shown, including budget models. Same workload for every model: 8,717 input and 935 output tokens per answer. Lower is better.
One answer is a typical full written reply to a general question (about 8,700 input and 935 output tokens across 165 real answers).
Budget models
Estimated US dollars per answer. Lower is better.
Flagship & mid-tier models
Estimated US dollars per answer. Lower is better.
Praus scores are self-reported. Praus is a compound system that can combine several models, tools and retrieval, and it is compared here with individual models. Scores come from samples of each test set, and the MBPP and TriviaQA results are from preliminary runs of about 50 questions. SimpleQA was run with web search on. Competitor scores were compiled on September 28, 2026. The cost figure for Praus is our estimated running cost; the other figures use public list prices. Methodology
An assistant for your everyday work.
An API for the software you build.
Talk to us about the app, API access or using Praus with your team.
We also built Lost Tribes. Explore Lost Tribes
