Google Opens Deep Review to Everyone With Two Free AI Document Checks a Week, but Its Accuracy Proof Rests on 29 Errors

Google Research opened Deep Review to the public on 14 October 2026. It is an agentic AI system that reads a long technical document, splits it into sections, checks each one in parallel and returns a report on logical gaps, math errors and claims that lack support.

Anyone can run two reviews a week for free. Heavier users can plug in their own Gemini API key, which moves the cost onto normal Gemini billing in USD. Google says uploaded files are never used to train its models but are stored for one year.

The pitch is strong: the tool already gave pre-submission feedback on more than 16,000 conference papers. The evidence behind the accuracy claims is thinner than the headline number suggests, so here is what is actually confirmed.

What Google launched

Deep Review is the public version of the Paper Assistant Tool, or PAT, an internal experiment Google ran with four computer science conferences over the past year. The Research@ 2026 roundup lists it as one of ten highlights from the event in Mountain View.

Research scientists Rajesh Jayaram and Vincent Cohen-Addad, who wrote the launch post, say the system has grown beyond academic papers. Google now pitches it for design documents, slide decks, legal agreements and financial reports.

The free portal lives at research.google/deepreview. Google also mentions a Deep Review API for teams that want to run checks inside their drafting or CI workflows, tied to a Gemini API key.

Free tier, API key and storage: the terms that matter

The public portal’s own configuration lists Gemini 4 Argon as the default model, with Gemini 3.8 Flash, Gemini 3.1 Pro preview and Gemini 3.5 Flash Lite as other options. Google has not said which of those the free tier always uses.

For API users, cost depends on model and document length. Gemini 4 Argon is priced at $2 per million input tokens and $10 per million output tokens until 30 November, then doubles to $4 and $20. Because Deep Review runs several agents per document, one review will use far more tokens than a single prompt. Google has not published a per-review cost estimate.

How the review pipeline works

Google’s argument is that one long prompt fails on long documents. The model runs out of reasoning budget, misses links between sections, or accepts subtle claims without checking. Deep Review uses four stages instead:

• Segmentation. A segmenter agent splits the document into logical parts such as introduction, methods and experiments. Segments can overlap but must cover everything.
• Adaptive budgeting. Dense sections, like proofs, get more compute than a related work section.
• Parallel verification. Separate reasoning agents check each segment while still seeing the full document for cross references.
• Synthesis and grounding. A final agent merges the reports, reruns consistency checks and uses external search to catch hallucinated citations or theorems.

That last step matters. In the June 2026 PAT preprint, the authors say the synthesis agent uses Google Search to check for things like papers or theorems that do not exist.

The pilot numbers, and what they really measure

Google’s figure of 16,000+ manuscripts comes from four opt-in pilots. Adding up Google’s own table gives about 16,220 papers: 220 at STOC 2026, 4,500+ at ICML 2026, 5,500 at NeurIPS 2026 and 6,000+ for ICLR 2027.

These are satisfaction surveys of authors who chose to try the tool. They show people liked it. They do not show how many real errors it caught or how often it was wrong.

The harder evidence is in the preprint. On a subset of the SPOT benchmark of retracted papers, PAT running on Gemini 3.1 Pro detected 89.7% of known equation and proof errors, against 55.2% for the same model zero-shot. That subset is 26 papers with 29 errors, graded by an automated LLM grader. It is a useful signal, but it is a small test of an older system, not of the public Argon setup.

The preprint is also candid about failure modes. Authors reported date hallucinations, PDF parsing problems, and the tool falsely claiming a proof was wrong. Google says it fixed the first two with better search and parsing. The third, it admits, is a problem for any LLM-based reviewer.

Confirmed vs unconfirmed

Why the timing makes sense

Google frames this as a “validation bottleneck.” It says ICLR, NeurIPS and ICML received over 74,000 submissions this year, up from about 45,000 in 2025, 32,000 in 2024 and 23,000 in 2023. That is more than three times the 2023 load in three years, with roughly the same pool of human reviewers.

Google also shipped a related pattern for developers. Its Antigravity coding tool gets a “Teamwork Deep Document Review Pattern,” invoked with the /teamwork-preview command, which applies similar multi-agent review to papers, drafts and technical proposals.

This fits a wider push from Google into agent tools. Its Gemini agent for planning work across Workspace and Microsoft 365 is still in private preview. Deep Review, by contrast, is open to anyone with a document today.

What it means for Indian developers

For students and researchers in India who submit to NeurIPS, ICML or ICLR, two free deep checks a week is real value. Google itself notes that detailed pre-submission critique has mostly depended on access to senior faculty or large labs. A free second reader that flags a broken proof or a missing baseline narrows that gap.

Engineers can use it on design docs and RFCs before review meetings. If you want it in CI, the API route bills through your Gemini key in USD, so set a budget alert first, since agentic runs can burn tokens quickly.

Be careful with client work. Contracts, financial statements or documents with personal data will sit on Google’s servers for a year. Teams handling regulated data under India’s DPDP Act, or under client NDAs, should get sign-off before uploading. If you work with Google’s open models on device instead, our look at EmbeddingGemma 2 covers a local option for search over your own files.

The bottom line

Deep Review is a free, well-designed second reader with a sensible pipeline and honest documentation of its weak spots. Use it to find things to check, not as a verdict. A tool that can wrongly call a correct proof broken still needs a human who knows the subject to have the last word.

FAQ

Is Google Deep Review free?

Yes, every user gets two document reviews per week at no cost. For more, you connect your own Gemini API key and pay normal Gemini usage rates. Google has not listed a separate Deep Review fee.

What file types does Deep Review accept?

Google’s launch post says PDF and plain text are supported first, with multi-file input coming soon. Some coverage lists more formats, so check the portal before you rely on Word or slide uploads.

Does Google train its AI on documents I upload?

Google says uploaded documents are never used to train, fine-tune or improve its models. They are stored for one year so you can see results in the Deep Review interface.

How accurate is Deep Review?

Google has not published accuracy figures for the public version. The earlier PAT system caught 89.7% of 29 known proof and equation errors in a small benchmark, and authors reported that it sometimes wrongly flagged correct proofs.

Can I use Deep Review for legal or financial documents?

Google says yes and suggests it for checking clauses and data tables. It has not shared any benchmark for those domains, and you should think about confidentiality before uploading, since files are kept for a year.