Anara 3.0: A new kind of scientific instrument

A few months ago we started experimenting with ways to direct AI agents to complete research tasks end-to-end. Instead of simply working along side us summarizing and editing, we wanted a system that could handle research work autonomously across the full literature synthesis workflow.

Today, we're introducing Anara 3.0, a major update to our model harness and research workspace. We rebuilt Anara from the ground up to handle long-running research tasks, reviewing up to 100,000 files at once and with superhuman accuracy.

Give it a goal like completing a literature review or mapping the competitive landscape for a drug class and Anara will search your library, the web and academic databases, synthesize what it finds to return a cited, verifiable document you can edit.

We're rolling out access today and through the rest of this week.

What's new in Anara 3.0

Agent sidebar

Agent now lives in a new right sidebar; you can access it by clicking the “Agent” button in the top right corner of the app. Describe what you want and Anara will get to work. The most effective prompts tell Anara what to look at and what you want back. But don’t worry too much about the perfect prompt because Anara will ask follow-up questions.

You'll see it working in the chat thread, communicating which files it's reading and tasks it's carrying out. Once complete, it produces a final response either in the chat window or if it creates something it will be open as a new tab in the workspace. You can expand each section to understand how it reached a specific conclusion.

The agent sidebar stays open as you move around the app, and your chats are saved in a new "Chats" section in the left sidebar. By default Agent draws context from your library, sources from the web and the files and folders you have in view. You can disable library and web as sources with the buttons at the bottom of the input.

97% more accurate than general purpose AI

We evaluated Anara 3.0 against six general-purpose AI platforms on a multi-document question-answering benchmark (MADQA). Each platform was tested under identical conditions: same documents, same questions, same scoring criteria.

Anara 3.0 answers more questions correctly than every other platform tested, including ChatGPT and NotebookLM.

Anara 3.0 scores 97% higher than the average of all other platforms tested (43%) at locating the exact page containing the answer.

500M+ papers indexed

We expanded the number of academic papers we have indexed from the open web now amounting to 500M+ across sources like arXiv and JSTOR. Life sciences sources are now built in as connectors @-mention PubMed, bioRxiv, medRxiv, or ClinicalTrials.gov directly in chat to search them natively.

A more flexible interface

The agent works across many files, folders and notes in a single session, and until now that meant a lot of jumping between pages. To solve this problem we introduced a tabs feature. You can now open any file or folder as a tab and drag and split to work on files in parallel. Agent will always have context of the tabs in view.

Ask chat

The old way of chatting with files and folders hasn't gone anywhere. We just call it “Ask” chat now. Use it to keep a separate, ongoing conversation scoped to that specific file or folder. You can return to these chats just as you have previously by clicking the dropdown at the top of the chat thread.

Where people and AI do research work together

This won't be the last time the interface for doing science changes. More powerful models will unlock new interaction patterns. We're looking forward to continue building, simplifying and transforming Anara to be the place where people and AI do research work together.

These features were crafted by us but driven by conversations with you. Don't hesitate to reach out with thoughts and comments. We read every piece of feedback we get.