How Praus compares.
Everything the frontier assistants do, on open models you can inspect — with the choices about your own work written down rather than assumed. Scan the whole picture before you ask us anything.
| Compared on | Praus | ChatGPT | Claude | Gemini |
|---|---|---|---|---|
| Models and intelligence | ||||
| Open-weight modelsModels anyone can inspect, not a sealed box. | Yes | No | No | No |
| Several models in one assistantNo single model has to be good at everything. | Yes | No | No | No |
| Routed for each questionThe model is chosen for the task, not by you. | Yes | No | No | No |
| Models can work togetherA harder question can draw on more than one. | Yes | No | No | No |
| Answers checked before you see themA verification pass looks for inconsistencies. | Yes | No | No | No |
| New models added as they shipBetter open models join without you switching. | Yes | No | No | No |
| Frontier-level benchmark scoresLeads on SimpleQA and GraphWalks, close on math. | Yes | Yes | Yes | Yes |
| What you can do | ||||
| Chat with follow-upsAsk, refine and keep going in one thread. | Yes | Yes | Yes | Yes |
| Web research with sourcesQuick answers that carry something to check. | Yes | Yes | Yes | Yes |
| Deep Research reportsLonger investigations, returned as a written brief. | Yes | Yes | Yes | Yes |
| Writing and draftingLetters, posts, summaries and long-form drafts. | Yes | Yes | Yes | Yes |
| Coding helpFix a bug or plan a build, with the reasoning shown. | Yes | Yes | Yes | Yes |
| ProjectsChats, files and notes for one piece of work. | Yes | Yes | Yes | Yes |
| Notes kept beside your chatsWrite and edit notes the assistant can work with. | Yes | No | No | No |
| Group chatsBring people into one conversation with the assistant. | Yes | Yes | No | No |
| Memory across your workIt remembers what you have told it, if you want. | Yes | Yes | Yes | Yes |
| Custom instructionsSet how answers should be written, once. | Yes | Yes | Yes | Yes |
| Quick or considered answersAuto for speed, Think for a complex question. | Yes | Yes | Yes | Yes |
| Files, images and voice | ||||
| Attach documentsBring a file into the chat and ask about it. | Yes | Yes | Yes | Yes |
| Reads images and scansPhotographs, screenshots and scanned pages. | Yes | Yes | Yes | Yes |
| DictationSpeak instead of typing. | Yes | Yes | Yes | Yes |
| Answers read aloudListen to a long answer instead of reading it. | Yes | Yes | Yes | Yes |
| iOS appThe same workspace on your phone. | Yes | Yes | Yes | Yes |
| Agents and developers | ||||
| Agent tasksMulti-step work carried out for you in the background. | Preview | Yes | Yes | Yes |
| Agent results verifiedSaved results are checked before they come back. | Yes | No | No | No |
| Developer APIThe same routed answers, called from your own server. | Yes | Yes | Yes | Yes |
| No model to choose per callRouting picks the model, so your code never has to. | Yes | No | No | No |
| Streaming responsesShow the answer as it is written, token by token. | Yes | Yes | Yes | Yes |
| Conversation contextSend earlier turns so the answer stays in context. | Yes | Yes | Yes | Yes |
| Organization API keysIssue and rotate keys for a whole team. | Yes | Yes | Yes | Yes |
| How access startsWe turn API access on per account, after a conversation. | On request | Self-serve | Self-serve | Self-serve |
| Data, privacy and sovereignty | ||||
| Training on your conversationsOpt-in only. Accepting the terms is not consent. | Opt-in | Opt-out | Opt-out | Opt-out |
| Temporary chats, never trained onA conversation kept out of history and training. | Yes | Yes | No | Yes |
| Minors’ content never trained onContent on profiles marked as a minor is excluded. | Yes | No | No | No |
| Identifiers masked before inferenceEmails, phones and SSNs swapped for placeholders. | Yes | No | No | No |
| Inference providers namedThe policy names who processes your text. | Yes | No | No | No |
| API content never trained onCommercial content is excluded under the API terms. | Yes | Yes | Yes | Yes |
| No sale of personal informationYour conversations are not a product for sale. | Yes | Yes | Yes | Yes |
| No advertising use of your chatsNothing you write is used to target ads. | Yes | Yes | Yes | Yes |
| Delete your account yourselfA delete control in Settings, not a support ticket. | Yes | Yes | Yes | Yes |
| Access, correction and exportAsk for a copy, a correction or a deletion. | Yes | Yes | Yes | Yes |
| Data residencyProcessing locations committed in writing. | By agreement | Enterprise only | Enterprise only | Enterprise only |
| Private deploymentRun it inside your own environment, by agreement. | By agreement | Via cloud partner | Via cloud partner | Via cloud partner |
| Published security contactA named address for reporting a vulnerability. | Yes | Yes | Yes | Yes |
| Accounts, plans and cost | ||||
| Free to startA free tier, before you decide to pay for anything. | Yes | Yes | Yes | Yes |
| Two adults, one household billSeparate private accounts on a single subscription. | Yes | No | No | No |
| Private from the plan payerPaying does not grant access to the other account. | Yes | No | No | No |
| Cost per answerEstimated running cost of one typical answer. | $0.00069 | $0.13392 | $0.05357 | $0.02865 |
| Public, reproducible testsNine public benchmarks you can run again yourself. | Yes | Vendor-reported | Vendor-reported | Vendor-reported |
Consumer assistants, as described in each provider’s published documentation in November 2026; business, enterprise and API tiers differ, and settings change. Cost per answer is the estimated running cost of a typical answer at published rates, comparing Praus with GPT-6 Astra, Claude Opus 5.5 and Gemini 3.1 Pro — the full list is under Benchmarks, and Praus results are at Hugging Face. Agents are a preview. Sovereign does not mean that every request stays on one device or inside one country: processing locations, safeguards and private deployment are set out in our Privacy Policy and agreed in writing for your organization.