Reviewer3 and Previum do the same job: they read your manuscript and return the objections a reviewer would raise before the journal does. There is no “they don’t really compete” here: they compete. So the comparison cannot run on adjectives; it runs on figures, and the figures come from their benchmark, not ours.
Reviewer3 publishes ReviewBench, an open benchmark with 1,000 eLife articles and the comments of the human reviewers, of Reviewer3 and of two general-purpose models, with an automatic judge and its code. We ran our engine through 200 of those articles with their code and their judge, and put the averages side by side. On agreement with human reviewers there is no significant difference; on anchoring each objection in the text and on saying why it matters, we come out ahead; on pointing at a specific claim of the paper, they do.
I write this comparison as an interested party, cards on the table: what Reviewer3 does better, what mine does better, what each costs and in which concrete case each one is the right call. Every third-party figure has its source one click away.
Reviewer3 has the best upload screen on the market and it deserves saying. You drop the PDF without creating an account, and forty seconds later you already have the paper’s central claims, the figures that do not add up and the references; the full cards arrive in under two minutes, with their severity and the button that takes you to the exact PDF page. A chat alongside lets you ask about each finding. It is fast, clean and asks for nothing until you want to see the rest.
On their benchmark there is also one metric where they beat us and we will not hide it: their comments point at a specific claim of the paper 77.1% of the time, ours 73.1%. Four points, with p < .001 over 200 articles. Part of that gap is that our full report also objects to method and reporting (guidelines, preregistration, declarations), which do not always hang from a claim; but the metric is theirs, the code is theirs and the result is that.
They also run three integrity checks that do not exist here, and you should know it before choosing: AI-generated text detection, retracted-citation warnings and self-citation counts. If your journal looks at those, or you are an editor screening at volume, that block of theirs has no equivalent in Previum.
And scale: if you run many manuscripts a month, $19 of subscription with unlimited reviews beats €12 per full report once you go past two or three. For a group reviewing daily, that is their ground.
ReviewBench, the benchmark and the code (GitHub, MIT) · Reviewer3 pricing and plans · Our run through their benchmark, with the JSON · Our benchmark against Reviewer3, step by step
Where we differ most is the anchor: 98.1% of our comments carry a concrete reference to the text (a table, a figure, a quoted sentence) versus 83.0% of theirs. It is not a matter of style: an objection without an anchor forces the author to hunt for where the problem is, and one with an anchor opens on the exact sentence. It is also what makes it possible to return your Word file with every comment in the margin, which is the deliverable you forward to a supervisor or a co-author.
And on two measures of substance: 97.2% of our comments that point at a claim could undermine it (89.7% of theirs), and 100% say why it matters (97.4%). On coverage of what human reviewers address and on stance agreement the figures are almost identical and the difference is not significant: on what an editor would call “agreeing with the real review”, it is a draw.
Outside the benchmark there are three things Reviewer3 does not do that go into the Basic report here: recomputing every p value from its degrees of freedom and flagging when the recomputed value changes the conclusion; checking APA 7 statistical reporting with rules; and returning the report in Spanish, Portuguese or French even when the manuscript is in English. And reference verification is published citation by citation on the 636 of the original Walters and Wilder benchmark, not on an unitemised subset.
Results on ReviewBench and on the citation benchmark, with intervals and JSON · Walters and Wilder (2023), the fabricated-citation benchmark, in Scientific Reports
Both of us check the bibliography, and both publish a percentage; but they do not measure the same thing. Reviewer3 publishes 98.5% accuracy on 476 citations from a set it does not itemize. We ran all 636 citations of the original Walters and Wilder benchmark (441 real and 195 fabricated by two language models) and publish the result citation by citation: 92.3% of the fabricated ones caught, 86.4% of the real ones recognized, 88.2% accuracy overall.
We lose on “real ones recognized” and it needs saying why: 60 of the 441 real citations carry some wrong detail (authors, title, year or journal), and our verifier requires two databases to agree on title and authors, so it flags them. For the benchmark that counts as a miss; for the author it is a useful warning, because the citation is wrong even though the work exists. Books are our genuine weak spot (80.9% versus 94.3% on articles) and it is on the list.
What does help you decide: in Previum the reference check is part of the Basic report, with your account, up to 30 references per pass (60 in the full report), and each one comes out with its verdict: confirmed in two or more databases, in a single one, with the DOI pointing to another work, with a DOI no database recognizes, or with no trace. “No trace” does not mean fabricated, and the report says so.
The 636 citations with their label and our verdict (JSON) · How a reference is verified, step by step
Reviewer3 shows the first card and the references for free; the rest of the report opens for $11 once or $19 a month. Previum shows for free the finding that decides, with their quote and their fix, plus the references, the recomputed statistics and the figures that contradict each other; the full report (every objection from two simulated reviewers, the editor’s letter, the action plan, the journals where it fits and the second round with the corrected version) costs €12, one-off.
The arithmetic is simple: one manuscript, €12 versus $11. Multiply by how many you run a month and compare with their $19: once you go past two or three, Reviewer3’s subscription is the better deal. If you review one paper a quarter, the annual difference is an order of magnitude in favor of the one-off; if your group runs manuscripts daily, look at their subscription or ask for an institutional license.
A caveat: I check their prices in September 2026 on their website and they change; mine have changed this year too. They are here to compare orders of magnitude, not to make the transfer.
Reviewer3 pricing · What the Basic report includes and what the full one adds, on the sample report
Both encrypt the manuscript and both say it is not used to train models. The difference is in the rest of the policy: Reviewer3’s declares targeted advertising and third-party pixels; on this site there is no advertising, the manuscript is not stored, not even with an account (the review is), and the report does not go out by email: it stays in your workspace and you share it yourself, with an encrypted link that expires.
On the criteria: Reviewer3 is a product of Reviewer3, Inc., and its backing is its benchmarks and the testimonials from professors it publishes on its front page. Previum is signed by a reviewer for Q1 journals in psychology and health sciences, and its catalog of objections comes from real open peer-review reports, with the rates per design published. That does not make it better on a benchmark; it makes it explainable when an objection does not sit right with you.
Reviewer3 privacy policy · Security and manuscript privacy on this site
Give the same manuscript to both. Each has a free layer that is enough to compare: the first card and the references on Reviewer3; the finding that decides, the references, the statistics and the figures on Previum. Look at three things: whether the main objection matches, whether it tells you where it is and what to do, and in which language you would rather read it.
If you do not want to upload anything yet, there are six sample reports on real, already published papers, one per discipline, generated with the engine as it is and not edited: nursing, physiotherapy, education, sports science, dentistry and psychology. You see the manuscript with every critique in its sentence and the checks underneath.
Previum does the same as Reviewer3 (reads the manuscript and returns a reviewer’s objections, with the quote and the fix) and delivers the report in Spanish, English, Portuguese or French, even if the manuscript is in another language. It also returns your Word file with every comment in the margin, recomputes the p values and checks the references against eight databases in the Basic report.
In September 2026, Reviewer3 shows the first card and the references for free, and the rest costs $11 per review or $19 a month. Previum has a Basic report, free with your account, and a full report at €12, one-off per manuscript, second round included. From the fourth manuscript a month on, their subscription is the better deal.
On their own benchmark, ReviewBench: 200 eLife articles, their code and their judge. On agreement with human reviewers there is no significant difference (93.4% versus 95.6% of the claims the humans address; 67.9% versus 70.4% with the same stance). They win on pointing at a specific claim (77.1% versus 73.1%); we win on anchoring in the text (98.1% versus 83.0%), on the objection being able to undermine the claim (97.2% versus 89.7%) and on saying why it matters (100% versus 97.4%). The JSON with intervals is on /en/evidence.
They publish 98.5% on 476 citations they do not itemize; we publish 88.2% on the 636 of the original benchmark, one by one. Most of our gap is real citations with some wrong detail (60 of 441), which our verifier flags because it requires two databases to agree on title and authors. Books are our weak spot. What each reference does share in Previum is its verdict and the databases that confirm it.
On Reviewer3 you upload without an account and create one afterwards to see the result. On Previum you sign in first, with Google, Microsoft, ORCID or an email link, and any real email works; the report stays in your account, with the Word and the PDF, and never goes out by email.
Yes: there are six sample reports on real, already published papers, one per discipline, generated with the engine without editing anything, with the manuscript and every critique marked in its sentence. They start at /en/ai-paper-reviewer/example.