Deep Review
Google Research · Deep Review: Google Research's Multi-Agent Reviewer for Papers and Technical Documents
Deep Review is a free-to-try Google Research tool that runs a multi-agent Gemini pipeline over a paper or technical document and returns a written review of logical gaps, mathematical errors, unsupported claims, and missing related work. It suits researchers checking a manuscript before submission and anyone who needs a rigorous second read of a dense proposal, design doc, or legal or financial document.
Best for
Researchers and graduate students who want a rigorous pre-submission check of a math- or experiment-heavy paper, and engineers, analysts, or lawyers who want a second read of a long technical, legal, or financial document
Not ideal for
Finding or summarizing literature, teams that need to review many documents a week without paying for Gemini API usage, documents that cannot leave the organization or be stored by Google, and workflows that need cross-document review today
Who it's for
Academic researchers, graduate students, and technical professionals who write or review long, dense documents such as papers, proposals, design docs, and legal or financial documents
Capabilities
- Four-stage agentic pipeline: a segmenter agent splits the document into sections, compute is budgeted by information density, specialist agents review sections in parallel with full-document context, and a synthesis agent reconciles the findings
- Checks logical consistency and mathematical statements, grounds factual claims, compares the work with existing literature, and suggests improvements
- Final cross-section consistency pass with external grounding checks and tool calls intended to catch hallucinated findings
- Four review depths: Low (single pass for early drafts), Medium, High (multi-agent, conference-level), and Extra High (exhaustive multi-pass for complex theoretical and empirical work)
- Optional custom context that steers the review, such as sections to focus on or the target conference, venue, or submission track
- Accepts PDF, TXT, Markdown, LaTeX (.tex), DOCX, PPTX, and HTML uploads up to 20 MB, or pasted text, and the FAQ lists Google Docs and Google Slides support
- Two free reviews per week, with a bring-your-own Gemini API key for more reviews, a choice of Gemini model, and skipping the queue
- Deep Review REST API to upload documents (or import them from Google Drive), start asynchronous reviews, and fetch a Markdown review plus up to 300 comments anchored to locations in the document, each with a severity level and an optional suggested edit
- API output includes per-reference citation checks that compare bibliography entries with OpenAlex and Crossref records and Google Search results to flag mismatched or likely hallucinated references
- Document history in the web app, with deletion of a document and all its reviews
Limitations
- Only two free reviews per week; more reviews require your own Gemini API key, billed at standard Gemini API rates (Google estimates about $3 to $20 for a 40-page document)
- Requires signing in with a Google Account
- One document per review: multi-file review across related documents, spreadsheets, and exhibits is announced as coming soon
- Uploads are limited to 20 MB, and password-protected or encrypted PDFs are rejected
- Reviews are asynchronous and typically take 10 to 30 minutes once started, and free reviews can wait in a queue under high demand (an API key skips the queue)
- Documents are stored by Google for up to a year for display, and the review may run web searches, which may not suit confidential, embargoed, or client documents
- No self-hosted or on-premises option is documented, and no team workspace or admin controls are described
- Google frames it as help with the most burdensome parts of review, with higher-level judgments left to humans, so it is not a substitute for expert peer review
Use cases
- Stress-testing a conference or journal manuscript for proof errors, unsupported claims, and missing related work before submission
- Getting a structured critique of a grant proposal, thesis chapter, or technical report while drafting
- Reviewing a software design document for unstated assumptions and unhandled edge cases
- Checking a contract or financial report for inconsistencies between clauses and data tables and for ambiguous terms
- Checking that a slide deck's takeaways match its charts and that the argument flows
- Running reviews from a script or CI pipeline through the Deep Review API with your own Gemini API key
Our take
Deep Review is unusual among research tools because it reviews the document you wrote rather than finding or summarizing other people's work, and it comes from a program that already ran pre-submission feedback at major CS conferences. For a researcher, two free reviews a week are enough to check a draft and then its revision, and the four depth levels plus a custom-context box let you aim a deep pass at the proofs or experiments that matter. Past the free tier the economics are transparent: you pay Gemini API rates per review, which on Google's own estimate can reach about $20 for a long paper at the deepest setting, so heavy use adds up. The main caution is confidentiality: documents are stored by Google for up to a year and reviews may search the web, so check venue, employer, or client rules before uploading unpublished or sensitive work, and treat the output as a checklist to verify rather than a verdict.
Who should use it
Researchers preparing conference or journal submissions, especially math-heavy or experimental papers, graduate students without access to many senior reviewers, and engineers, analysts, or lawyers who want a fast structured critique of a long design doc, proposal, contract, or report.
Who should skip it
People looking for a literature search or Q&A tool, anyone whose documents cannot be stored by a third party or exposed to web-search lookups, teams that need many reviews a week without Gemini API spend, and anyone hoping to replace human peer review rather than prepare for it.
Strengths
- Reviews the document you wrote instead of searching literature, including proofs, claims, and cross-section consistency
- Grew out of a pre-submission feedback program run with STOC, ICML, NeurIPS, and ICLR
- Two free reviews a week, then pay-as-you-go at standard Gemini API rates with no added Deep Review fee
- Accepts common formats (PDF, LaTeX, DOCX, PPTX, Markdown, HTML, text) and has a REST API
- Documents are not used for model training unless you opt in
Weaknesses
- Free use is capped at two reviews a week
- Reviews take 10 to 30 minutes and can queue under high demand
- Single-document only for now
- Documents are stored by Google for up to a year and reviews may run web searches
- No documented team, admin, or self-hosted options
Deep Review pricing
Free
$0
- 2 reviews per week
- Gemini 4 Argon at Medium depth
- Google Account sign-in required
With your Gemini API key
Gemini API usage
- Reviews beyond the weekly free quota
- Billed by Google at standard Gemini API rates, no added Deep Review fee
- Skips the queue under high demand
- Programmatic access through the Deep Review API
Free tier limits: Two document reviews per week per user.
Note: Google estimates a 40-page document costs about $3 on Low depth with Gemini 3.8 Flash and about $20 with Gemini 4 Argon on Extra High, and says actual charges vary by document.
Technical specs
Modalities
Text
API pricing
Bring your own Gemini API key: each review's Gemini API calls are billed at standard Gemini API rates, with no additional Deep Review charge
Available models
Where Deep Review excels
Pre-submission check of a theory or ML paper
A High or Extra High review focuses extra compute on dense proofs and experimental sections, with custom context naming the target venue, so authors can fix errors and gaps before reviewers see them.
Second read of an engineering design doc
Google suggests Deep Review for catching unstated architectural assumptions and edge cases, which helps when a design is too long or specialized to get a careful human review quickly.
Consistency check of a contract or financial report
The review looks for contradictions between clauses and data tables and for ambiguous terms, the kind of internal inconsistency that is tedious to find by hand across a long document.
How Deep Review fits alongside other tools
Deep Review and Gemini Notebook (formerly NotebookLM)
Gemini Notebook (formerly NotebookLM), also from Google, works over sources you add: it answers questions, summarizes, and produces study material such as audio and video overviews. Deep Review reads one document critically and returns a review of its errors and weak spots. A writer might use Gemini Notebook to work through background material and Deep Review to check the resulting draft.
Deep Review and Scite
Scite classifies how papers cite each other, as supporting, contrasting, or mentioning, which helps check whether the work you cite still stands. Deep Review checks the logic, math, and claims inside your own manuscript and compares it with related work. Scite covers the citation side of a pre-submission check, and Deep Review covers the content.
Deep Review and Elicit
Elicit searches and extracts data from the research literature for reviews and evidence synthesis. Deep Review does not search for papers to read: it critiques a document you upload. Elicit helps before and during writing, and Deep Review helps once there is a draft to check.
Frequently asked questions
What is Google Deep Review?
Deep Review is a Google Research tool that reviews a paper or technical document with a multi-agent Gemini pipeline. It splits the document into sections, has specialist agents check each section in parallel for logical, mathematical, and factual problems, compares the work with existing literature, and combines the results into one review. It is the public successor to Google's Paper Assistant Tool (PAT).
Is Deep Review free?
Every user gets two free reviews per week, which run Gemini 4 Argon at Medium depth. For more, you attach your own Gemini API key and pay only the Gemini API calls the review makes, at standard rates. Google says there is no additional charge for Deep Review itself and estimates about $3 (Low depth with Gemini 3.8 Flash) to $20 (Gemini 4 Argon on Extra High) for a 40-page document, with actual costs varying by document.
What file types does Deep Review accept?
The upload form accepts PDF, TXT, Markdown, LaTeX (.tex), DOCX, PPTX, and HTML files up to 20 MB, and you can also paste text. The FAQ also lists Google Docs and Google Slides, and the API can import a document from Google Drive. Password-protected PDFs are rejected, and reviewing several related files together is announced as coming soon.
Does Google use my documents to train models?
Google says uploaded documents are not used to train, fine-tune, or improve Google models, though you can opt in to let the research team use your data to improve Deep Review. Documents are stored for up to one year so you can view them and their reviews in the app, and the review may perform web searches to find related work. You can delete a document and its reviews.
Does Deep Review have an API?
Yes. The Deep Review API lets you upload a document or import one from Google Drive, start an asynchronous review with a chosen depth, model, and custom instructions, and retrieve a Markdown review plus up to 300 comments anchored to locations in the document. Google positions it for use with your Gemini API key, which is billed for the review's model calls.
Can Deep Review review documents other than research papers?
Yes. It started with computer science papers, but Google now suggests it for design documents, presentations, and legal and financial documents, where it checks internal consistency, flags ambiguous terms, and grounds claims against supporting evidence.
How long does a Deep Review take, and which depth should I choose?
Reviews run asynchronously and typically take 10 to 30 minutes once started, longer when free reviews are queued under high demand. Low is a quick single-pass check for early drafts, Medium balances speed and depth, High is a multi-agent, conference-level evaluation, and Extra High is an exhaustive multi-pass review for complex theoretical and empirical work. Free reviews run at Medium.
Deep Review vs Gemini Notebook (NotebookLM): what is the difference?
Gemini Notebook, formerly NotebookLM, answers questions about, summarizes, and builds study material from the sources you add. Deep Review critiques a single document: it looks for errors, unsupported claims, and inconsistencies, and returns a review rather than a conversation.
Integrations & fit
About Deep Review
Deep Review grew out of Google's Paper Assistant Tool (PAT), which gave opted-in authors private pre-submission feedback in pilots with STOC, ICML, NeurIPS, and ICLR; Google's launch post puts that at more than 16,000 manuscripts. Since October 14, 2026 it has been publicly available worldwide; you sign in with a Google Account to use it. You upload a document or paste text, optionally add context such as sections to focus on or the target venue, pick a review depth, and get a review back, typically 10 to 30 minutes after it starts running. Under the hood, a segmenter agent splits the document into logical sections, the pipeline allocates more compute to dense parts such as proofs, specialist agents review each section in parallel with the full document in context, and a synthesis agent checks consistency across sections and runs external grounding checks and tool calls as a guard against hallucinated findings. The point is critique of a document you already have, not finding literature or answering questions about sources, and Google says it leaves higher-level judgments to humans. Everyone gets two free reviews a week (Gemini 4 Argon at Medium depth); beyond that you attach your own Gemini API key and pay only standard Gemini API rates, which Google estimates at about $3 to $20 for a 40-page document depending on model and depth. The same key unlocks a REST API for running reviews from scripts or CI. Uploaded documents are kept for up to a year for display and are not used for model training unless you opt in, but they are processed by Google and the review may run web searches, so check confidentiality rules before uploading unpublished or client material. Separately, Google added a Deep Review-inspired document review pattern to its Antigravity agentic development platform.
Updates from Deep Review
Google Research released Deep Review, the successor to its Paper Assistant Tool, to users worldwide. Everyone gets two free reviews a week, a Gemini API key unlocks more reviews and the Deep Review API, and support for Gemini 4 Argon was added. Google also extended it beyond academic papers to documents such as design docs, presentations, and legal and financial documents.
Are you the founder? Claim this listing →