Why not just ask one model?
Any single model can be confidently wrong, and one answer never tells you which case you're in. On consens.io, other models check the work of the one that answers. They are independent, like an auditor: they never steer the agent, and it never sees their notes. Where they object, the sentence in the answer carries their mark.
What the benchmark measures
Our evaluation on 314 closed-book MMLU-Pro questions tests the consensus engine: several models answer, and one writes a combined answer from theirs. That engine still runs Watches, Topics and the API, but it is not how the app answers, so the benchmark says nothing about the reviewers. It is a limited evaluation of one configuration, and models can share the same error.
What it does
You ask a question or give a task. One AI, the chat model you pick, searches the web, reads sources and works with your files; for a task it can use a browser in its own workspace. Then it writes the answer. Other AI models check along: they follow every step and mark sentences they find wrong, doubtful or incomplete.
It is useful for research, learning, drafting, and everyday questions, but it is not a substitute for your own judgement.
How it works
- You choose the AI that does the work and the AIs that check it, or start from a preset: Daily, Open Models or High Quality.
- The backend sends requests through OpenRouter to model endpoints and may use web-search services.
- The reviewers follow a log of every step that the app writes, not the agent's own account of it. Before a step with an effect outside the chat, such as sending, uploading or typing into a website, the agent waits for their go.
- If a reviewer sees a safety problem or a hidden instruction, it pauses the run and you decide.
- When independent views help, the agent asks the reviewers' models the same question. Open Answers to read what each said, or compare two answers side by side.
Check a note, then keep the conversation
A reviewer's note is marked at the sentence it concerns: wrong, doubtful or incomplete. Open it to see who noted it, in what words and with which source. Unmarked text means no reviewer objected, not that it is confirmed.
Signed-in chats save automatically. Ask a follow-up in the same conversation, reopen it from the sidebar, and optionally use Memory to carry your preferences into later chats. Memory can be reviewed, edited and switched off.
From a private chat to a public answer
Publishing is a separate action. A shared answer lets readers inspect it, the models involved and its sources without an account. Explore published examples in the Consensus Library.
Watches rerun a question on a schedule and track changes over time. Public Watches and Topics show dated versions and how model positions move; readers can follow updates by email.
Integrated providers
consens.io is built around major model providers and supporting search tools. The exact line-up may change over time as models are added, updated, or removed.
Every model call goes to the fastest available provider endpoint that does not collect your prompts. Endpoints may keep requests for a limited period under their own terms, and web search is a separate service; the privacy policy explains what the routing does and does not cover.
OpenAI
ChatGPT
General reasoning, writing, coding, and structured output.
Anthropic
Claude
Long-form analysis, careful summaries, and nuanced answers.
Gemini
Fast research workflows and multimodal-capable model families.
Mistral AI
Mistral
Efficient responses and practical European model options.
DeepSeek
DeepSeek
Strong reasoning and coding at a low cost.
xAI
Grok
Conversational answers with current, real-time context.
Moonshot AI
Kimi
Efficient long-context answers, research, and practical reasoning.
Z.ai
GLM
General-purpose reasoning, writing, coding, and multimodal workflows.
Meta
Muse
Long-context Ember for everyday runs, open-weight Glimmer as an alternative.
Alibaba
Qwen
Open-weight Qwen3.8 27B for everyday runs, the Max flagship for long, multimodal reasoning.
Xiaomi
MiMo
Open-weight MiMo V2.6 Flash for fast everyday runs, the larger V2.6 Pro for demanding questions.
Project and maintainer
consens.io is built and maintained by Max Plack in Germany. It grew from a workflow for comparing AI answers into an independent product. Operator and contact details are in the Imprint.
Something on consens.io doesn’t work for you, for example with a screen reader, the keyboard or zoom? Write to contact@consens.io and we’ll look into it.
What it costs
consens.io is currently free during early access. The product is in active development. Usage limits apply, and some features are available only to selected accounts.
Paid plans are coming. There is no subscription to buy yet. Any paid offer would require a separate express agreement; free use does not automatically become paid. Feedback and feature requests are welcome at contact@consens.io.
For more runs, connect your own OpenRouter API key in the app. Requests still pass through the consens.io server, and OpenRouter charges your account under its terms.
Data handling
- Prompts, enabled Memory and conversation context are used to call AI models and, when invoked, search services.
- Signed-in conversations and their sidebar bookmarks are saved automatically until you delete them or your account. Publishing a private conversation is a separate action.
- Login and account features are handled through authentication infrastructure.
- If you use your own API keys, they are stored locally in your browser and only pass through the server to reach the selected provider; they are never stored server-side.
- There are no embedded ad-network tracking pixels. We use Umami Cloud for usage analytics, as described in the Privacy Notice.
For full details, read the Privacy Notice. Avoid secrets, highly sensitive data, and information that normally requires professional advice.