Greptile

By Minel Gunesoglu, founder. I build every published score from dated public evidence and read the community record behind it.

Section 2.2 of Greptile's terms of service lets the vendor train its own AI models on de-identified customer material, and the same section hands the customer a switch that stops it. Read on 18 August 2026 off a document stamped December 2024, the permission runs: "use such data to train and improve artificial intelligence algorithms and models", followed by "Company will solely and exclusively own the De-Identified Data and AI Training and Learnings". The switch sits in the same clause. A customer who does not want this "may manage preferences and settings related to AI Training and Learnings accordingly within Customer's account on the Platform", and the document publishes [email protected] for questions about opting out.

Both halves are the finding, and a page that prints one without the other has described a different clause. The rest of this ledger is what a $30 seat actually buys once the credit meter is counted, what the agreement commits each side to, what fourteen solicited reviews can carry, and where the 66.2/100 comes from.

What Greptile is, and what it reads before it writes a comment

Greptile is an AI code-review product that comments on pull requests. Its own pricing page, at greptile.com/pricing, sells it per seat with unlimited repositories, and its G2 listing files it on a breadcrumb reading "DevSecOps Software > Secure Code Review Software > Greptile". The unit that decides the bill is neither the repository nor the seat. It is the review, and the pricing page divides reviews into two kinds at two rates: "One standard review uses 1 credit, one TREX review uses 3 credits, and additional credits are $1 each."

What this page does not do is describe how the product works inside. The record I hold for it is four documents (the pricing page, the terms of service, the G2 listing and one public benchmark) and not one of them documents retrieval, indexing or model choice. I am not filling that in from general knowledge. What is documented is the surface a buyer signs for: what it costs, what the meter counts, what the contract permits, and how it scored where somebody measured it.

What Greptile's terms permit: AI training on de-identified data, and the switch that stops it

The training permission is not buried in an appendix. It is a headed subsection, and it is written in two directions at once. On the vendor's side, section 2.2 promises a de-identification step: "Company will implement appropriate measures and controls designed to remove personally identifiable information, customer-specific references, and sensitive data from De-Identified Data". On the customer's side, once the preference is changed, "Company will thereafter cease further AI training under this Section 2.2".

Three things follow for a buyer, and only three. The default is permission, so a team that assumed otherwise assumed wrong. The remedy is an account setting rather than a negotiation, which makes it unusually cheap to act on: change it on day one and keep the confirmation. And this is a reading of what a document said on 18 August 2026, not a measurement of whether the vendor honours it. Nothing on this site claims to test that.

The agreement also draws no line between a paid seat and a free one. The phrases free plan, free tier and free version appear nowhere in it, which I searched for rather than assumed, so a solo developer on Starter is reading the same clause as a funded team.

Greptile pricing: the $30 seat is a floor, and the meter runs on reviews

Pro is $30 per seat per month. There is no annual view to convert it against: the pricing page carries no billing toggle at all, and a search of the rendered text for "pay monthly", "pay annually", "billed monthly" and "billed annually" returned nothing. The absence is worth stating, because on most vendor pages the headline figure is the annual one and the month-to-month rate is higher. Here there is a single number, read in a browser window on 18 August 2026.

That number is a floor rather than a price. A seat arrives with 50 credits, and credits are what a review is drawn from.

What Greptile credits buy, and what happens after the first 50

One standard review costs a credit and a TREX review costs three, so a seat covers fifty standard reviews in a month, or roughly sixteen of the heavier kind, and every review after that is a dollar. The meter counts reviews, not people. A four-person team that opens two hundred small pull requests a month runs past its allowance long before a two-person team that opens twenty large ones, and nothing on the invoice would have predicted it.

So the thing to model before signing is not headcount. It is merge rate multiplied by the review type you intend to use. A shop with a busy trunk and small commits should price the overage first and treat $30 as an entry fee.

Is Greptile free? The Starter tier, the trial and the two discounts

Yes, for one person. In the vendor's own words: "Starter is free for one active developer and includes unlimited repositories and 50 credits per month". Paid tiers open with a 14-day free trial, offered on the same page.

Two published discounts sit beside that, and both are conditional. "Greptile is free for qualified non-commercial projects with MIT or Apache licenses." And: "Pre-Series A startups with under $2M revenue in the past 12 months get 50% off Greptile." Neither is a standing consumer offer; both require you to qualify rather than to click.

The pricing change Greptile users were arguing about in August 2026

Greptile moved to a base-plus-usage structure recently enough that users were still objecting publicly in August 2026, on a Hacker News thread and in posts from named users on X, read on 18 August 2026. I have not captured those threads and I quote none of them, so take this paragraph as a dated record that the objection exists rather than as evidence of what was said inside it.

Two dates are worth holding side by side. The agreement governing the purchase is stamped December 2024. The pricing structure it governs changed after that. That is an observation about two dates. It is not a claim about anyone's intent, and no legal conclusion follows from it.

Is Greptile safe for a private repository? Indemnity, ownership and the fee cap

On the indemnity, the answer is better than this category's usual shape. Section 5.4 runs both ways: "Each Party will indemnify, defend, and hold harmless the other Party for any direct infringement caused by such indemnifying Party's intellectual property provided under this Agreement". On the tally this desk kept on 16 August 2026, nineteen of the twenty-five agreements read for this site pointed the indemnity one way, with the customer covering the vendor. A mutual clause is the exception rather than the rule.

It is also a bounded one. Section 5.5 caps each party's total liability at "the fees actually paid by the Customer to Company during the prior twelve (12) months" and excludes indirect and consequential damages. For a team on one seat, that ceiling is a few hundred dollars. A mutual indemnity under a twelve-month fee cap is a real position and a limited one, and both halves belong in the same sentence.

Who owns the code Greptile suggests, and why the answer is unclear

Your own material is settled. "All Customer Data will be considered proprietary to Customer", and the ownership clause reserves the rest to the vendor: "Company owns and reserves all other rights, title, and interest in and to the Platform and Services".

The gap sits between those two sentences. Customer Data is defined as what the customer provides. No clause in the agreement assigns the review suggestions the product generates, and the word Output does not appear in the document at all, the concept lives under Customer Data or nowhere. In practice a comment you accept becomes a commit in your repository, which is your material under the first clause. On paper, the moment before that is unaddressed. The card records this position as unclear rather than resolving it in either direction, and that is why the contract row is not higher than it is.

What a Greptile refund requires, and why the trial matters more

The default is blunt: "All fees and other amounts paid or payable by Customer under this Agreement are non-refundable and non-cancellable, except as otherwise expressly provided". The express exception is real, and it is in section 5.2, which turns up only if you read past the word except. If the vendor materially breaches and does not cure, the customer may terminate and receive back an equitable portion of the fees pre-paid for services not delivered.

That is a breach remedy, not a cooling-off window. Nobody is getting money back for changing their mind, and the 14-day trial is the practical mitigation: the non-refundable term bites after a trial rather than on first contact. Running the other way, an overdue account may be suspended without further notice and carries "any interest due at the rate of 18% per annum".

What Greptile's 14 G2 reviews are, and what an incentivized sample cannot tell you

Greptile's G2 rating is 3.8 out of 5 across 14 reviews. That figure came off the live page in a browser window on 18 August 2026, because the platform serves nothing to a script. Fourteen is thin, barely above the minimum sample this desk will score at all, and the thinnest crowd base behind any card in this category on 18 August 2026. It belongs beside the number rather than in a footnote.

The composition matters more than the count. Our scoring model requires the solicitation regime to be read off the platform's own words rather than inferred, and G2 states it outright, review after review: "Validated Reviewer | Incentivized | Source: G2 invite". Two of the rows carrying that marker are dated 13 August 2026 and 5 August 2026. One row also reads "Business partner of the seller".

This is therefore not a rating that arrived unprompted. It is fourteen answers from people who were invited and given something for answering, one of whom discloses a commercial relationship with the seller. A 3.8 built that way is a different object from a 3.8 that accumulated on its own, and a reader who treats the two as interchangeable has been misled by me rather than by G2. The cell is pulled toward the neutral middle for exactly that reason: usability publishes 71.9, where the unshrunk reading was 76.0.

Nothing else was opened. No second review platform carries a captured sample for this product in our record, so the crowd picture here is one page, one date, fourteen voices.

What Greptile's 66.2/100 rests on, cell by dated cell

Every cell above is a reading of a document or a published board on a stated day. None of it is a trial of the product: I have not put Greptile in front of a repository, and no sentence here should be read as though I had. The rules behind these four cells, and the limits stated openly beside them, live on the methodology page.

The crowd cell is the thinnest evidence on this card, and the card says so where the figure appears. Capability is the sturdiest, because that board publishes a sample size per tool and Greptile's is the third largest on it.

How accurate is Greptile? 58.7% F1 and 75.3% precision on a public board, read 15 August 2026

The capability cell reads a public online tracker that scores AI review bots on real open-source pull requests. Its method asks, in its own words, "did the bot's suggestions lead to real code changes?", and the table it heads is labelled an "Overall online leaderboard balancing noise and thoroughness". Greptile's row on 15 August 2026: 58.7% F1, 75.3% precision, 48.1% recall, over 447 pull requests.

The rank is the least interesting part of that row. Seventh of eleven on F1 sounds middling, and the field is tight (42.9% to 62.4%, with the middle at 59.2%) so seventh is a point and a half off the median rather than a different class of tool. What the F1 column flattens is the shape underneath. At 75.3%, Greptile's precision is second on the whole board, behind only Cursor at 75.6%, while its recall of 48.1% sits near the bottom. That is a reviewer tuned to be right rather than thorough: it comments less often and is correct more often when it does. F1 weights the two equally, which is precisely the view in which that trade disappears.

Whether the trade is the right one is a question about your team, not about the tool. If a noisy bot gets muted after a fortnight, precision is what you are buying. If the cost you fear is a bug nobody flagged, recall is, and 48.1% is a hard number to accept.

Two caveats travel with any figure from this board, and neither is decoration. The board is parameterised, a slider trades noise against thoroughness, and the order changes as it moves, so ours is stated as the balanced setting. It also moves month to month; its own chart shows tools changing places between 13 July and 10 August 2026. A number lifted from it means nothing without both the setting and the date, so this one carries both and expires on 15 September 2026.

Greptile alternatives, and what the other scored cards in this category read

Named, not ranked. Other products here carry a card built the same way, and the sensible use of them is to read what each one actually measured rather than to treat the distance between composites as a recommendation.

CodeRabbit reads 73.3/100, and its card records per-seat billing with no permanent free tier for private repositories. Qodo's profile reads 68.8/100, and its card records a credit pack pooled across a whole team rather than a per-seat rate, the closest structural relative to Greptile's meter, with the difference that Greptile charges for the seat and the meter while that one charges for the pool alone.

If the deciding factor is the training clause at the top of this page, read each agreement rather than each score. Those clauses vary far more between vendors in this category than the prices do. The full category ledger lists each scored card with its evidence dates.

Who Greptile suits, and who should not buy it on this evidence

Buy it if you want a reviewer that speaks less and is right more often, and if your merge rate is modest enough that fifty reviews a seat is a ceiling you rarely touch. A solo developer gets a genuinely free tier with unlimited repositories. An MIT or Apache non-commercial project may qualify for free use, and a pre-Series A company under $2M in revenue for half off. Those are real offers, published on the vendor's own page, and they are the strongest part of the value case.

Do not buy it if recall is what you are short of: 48.1% on the balanced view of a public board is the figure to argue with, and no discount changes it. Do not buy it if your repository cannot be used for model training under any configuration and you are unwilling to set the preference and record that you did, because the permission is the default and the opt-out is yours to exercise. Do not buy it if you open many small pull requests on a fixed budget, since the meter counts reviews and the seat price will understate the bill. And if a broad customer record is what you decide on, weight this card accordingly: fourteen invited reviews is a thin base to carry an opinion, and I would rather say that than dress it up.

Score card built and page written by Minel Gunesoglu, founder of Vouch. Paid submission buys review speed only, never a listing and never a score.