diffox is the cheapest of the top five by score: $0.41 per review
We compared 23 AI code review tools on Martian’s independent benchmark. diffox is the cheapest of the top five by score: $0.41 a review; the other four $1.50–5.63.
Summary. We took the results of Martian’s independent, open-source code review benchmark on its 50 pull requests and set them beside the monthly prices on each tool’s public pricing page. The charts show the price per review of 11 of the 21 competitors. For the other 10, we list below what their vendors publish and why we left them out of the charts. We read competitor prices on the vendors’ own pages on 9 October 2026; diffox’s price is what our billing records charged for these 50 reviews.
- diffox is the cheapest of the five highest-scoring tools. The top five are diffox ultra (66.5), Qodo Extended (65.1), Cubic v2 (64.9), Augment (64.0) and diffox (60.9). A diffox review costs $0.41; the other four charge $1.50–5.63 a review, 3.7 to 13.8 times as much. diffox ultra is first among 23 tools.
- The same score for a quarter of the price. Qodo v2 scores level with diffox (60.6) and charges $1.67 a review, 4 times diffox’s price.
- The lowest price for each real bug caught. Each real bug caught costs $0.19 with diffox and $0.69 with diffox ultra; the lowest figure among the competitors is $0.73.
- $73 a month for a 20-developer team. If every developer opens 9 pull requests a month, diffox’s bill is $73; for the competitors we compared it is $180 to $800 a month.
diffox is the cheapest of the top five
In the first post we measured quality on Martian’s independent benchmark: 50 pull requests, 158 known bugs, judged by Claude Opus 4.5. This post adds each tool’s price. Below, the tools are sorted from cheapest to most expensive. Each row carries the benchmark’s three quality measures: F2 (the ranking measure), precision (how many of the comments written are real bugs) and recall (how many of the 158 known bugs were caught).
Read the chart from the top down: however high the price goes, no tool scores above diffox ultra. A review on diffox cost $0.40–0.41 at all three settings of the time. The five highest-scoring tools and their price per review:
| Tool | F2 | Price per review | Times diffox’s price |
|---|---|---|---|
| diffox ultra · High sensitivity | 66.5 | $1.64 | 4.0× |
| Qodo Extended* | 65.1 | $1.67+ | at least 4.1× |
| Cubic v2 | 64.9 | $3.51–5.63 | 8.6–13.8× |
| Augment | 64.0 | $1.50 | 3.7× |
| diffox · High sensitivity | 60.9 | $0.41 | 1× |
The other 18 competitors score below diffox; all of them are in the tables below. * Qodo Extended is not for sale and has no price: Qodo describes it as “currently only research results and not yet available in production”. We show Qodo’s average per review instead, as a lower bound (+).
Qodo v2 scores level with diffox (60.6) and costs $1.67 a review, so diffox gives the same quality for a quarter of the price. Six of the seven subscription competitors score below diffox (40.5–53.6). Assuming 9 pull requests a month, they cost 4 to 8 times as much per pull request. The seventh, Cubic v2, scores higher but costs 8.6 to 13.8 times as much. diffox ultra has the benchmark’s highest score (66.5) and costs $1.64 a review. That puts it in the same price band as Qodo Extended (at least $1.67) and Augment ($1.50), whose scores come close to ultra’s; Cubic v2 costs 2.1 to 3.4 times as much.
Most competitors are sold as a monthly subscription per developer. We worked out their price per review by dividing the monthly fee by the number of pull requests a developer opens a month; that number varies from person to person, so we give the price as a range (the calculation and sources are below). For Qodo, Augment and Cursor Bugbot we used the average per review that the vendor itself publishes.
The lowest price per bug caught
The most direct way to put price and quality into one number is to divide the price of reviewing all 50 pull requests by the number of real bugs caught.
At the High setting of the time, one real bug caught cost $0.19 with diffox and $0.69 with diffox ultra. The lowest among the competitors is Augment ($0.73); among subscription tools the range is $0.95 (Sourcery) to $2.79 (Baz). Even on the assumption that favors subscription tools most (a developer opens 11 pull requests a month), the best is Sourcery at $0.76; diffox ultra, at $0.69, is lower still. This calculation leaves out false alarms; the next section shows them.
The price barely depends on how many findings get posted
We measured diffox and diffox ultra at three settings: High, Medium and Low. All three ran the same review; only the number of findings they posted on the pull request differed. With diffox the three cost almost the same: $0.40–0.41. High caught the most bugs. Low posted only the findings it was surest of, and of the three settings it had the highest share of comments that were real bugs.
At Low, both diffox and diffox ultra were more precise than every competitor: 76% and 78%, against 67% for the best competitor. At High, diffox ultra’s recall was ahead of every competitor’s: 75%, against 66% for the best. diffox at High (67%) was level with that best competitor. The three settings of the time are described in detail in the first post.
A 20-developer team pays $73 a month
We ran the same numbers for a team: 20 developers, each opening 9 pull requests a month, which makes 180 reviews a month. diffox, Qodo, Augment and Cursor Bugbot charge for each review; the others charge a fixed monthly fee per developer.
| Tool | F2 | Monthly bill |
|---|---|---|
| diffox · High sensitivity | 60.9 | $73 |
| Cursor Bugbot | 48.6 | $180–270 |
| Augment | 64.0 | $270 |
| diffox ultra · High sensitivity | 66.5 | $295 |
| Qodo Extended* | 65.1 | $300+ |
| Qodo v2 | 60.6 | $300 |
| Sourcery | 51.5 | $300 |
| Gitar | 53.6 | $500 |
| Greptile v4 | 51.2 | $600 |
| CodeRabbit | 51.1 | $600 |
| CodeAnt | 40.9 | $600 |
| Baz | 40.5 | $600 |
| Cubic v2 | 64.9 | $800 |
Calculation: 20 developers × 9 pull requests a month = 180 reviews a month; each pull request is reviewed once. For subscription tools, the monthly list price × 20. If developers open between 7 and 11 pull requests a month, diffox costs $58–93, diffox ultra $233–374 and Qodo $237–380; the subscription tools’ bills do not change. * Qodo Extended is not for sale and has no price: Qodo describes it as “currently only research results and not yet available in production”. We show Qodo’s average per review instead, as a lower bound (+).
The cheapest competitors cost $180–270 a month (Cursor Bugbot, score 48.6) and $270 (Augment, 64.0). Sourcery (51.5) costs $300 a month, Qodo v2 (60.6) $300 and Qodo Extended (65.1) at least $300. diffox scores 60.9 for $73 a month; diffox ultra reaches the benchmark’s highest score, 66.5, for $295.
Estimate your monthly review budget
Estimated monthly total
Developers × PRs per developer × reviews per PR × average review price. This is a scenario, not a quote: actual prices vary by PR size and repository.
How many pull requests a month before diffox stops being cheaper?
The price per review of a subscription tool depends on how many pull requests a developer opens a month. Rather than guess that number, we calculated the break-even point: the number of pull requests at which the tool’s monthly fee equals what diffox charges for the same number of reviews. To the left of the bar’s end, diffox is cheaper.
A typical developer opens about 9 pull requests a month (shaded area 7–11). At that volume diffox is cheaper than every subscription tool we compared: the lowest threshold is Sourcery’s, at 37 pull requests a month, about 4 times typical use. Greptile has no threshold: its monthly fee includes 50 reviews, and after that it charges $1 per credit; diffox is cheaper at every volume. For diffox ultra the threshold lies between 9 (Sourcery) and 24 (Cubic v2): for a typical developer ultra costs the same as Sourcery and less than the other subscription tools.
If a developer opens more than 37 pull requests a month, Sourcery is cheaper than diffox; if that is your volume, you can read the answer off the chart.
How many pull requests does a typical developer open a month?
We found no published per-person figure for pull requests opened; there is one for pull requests merged. In DX’s 2026 report the median for developers who use AI weekly is 1.9 merged pull requests a week (8.2 a month); it is 1.5 for those who do not use AI at all and 2.4 for daily users. In Graphite’s research the median for active developers is 2 a week (8.7 a month). The two agree. To go from merged to opened we used the data in GitHub’s 2025 Octoverse: 47.5 million pull requests are opened on the platform each month and 43.2 million are merged, a ratio of 1.10. From that, the typical value is 9 pull requests opened a month, with a range of 7 to 11.
Work out the prices yourself
A price comparison is only as good as its sources, so every figure here shows where it comes from.
- Every competitor price is on the vendor’s own page. We took the list price on monthly billing, or the average per review that the vendor publishes; annual discounts are not included. We read the pages on 9 October 2026; each row’s source is linked in the table below.
- diffox’s price is written under every review. The last line of the summary comment shows that review’s charge; on your own pull requests you see the real price.
- The calculation is one line. For a subscription tool, the price per pull request is the monthly fee divided by the number of pull requests opened a month:
Sourcery $15 ÷ 9.1 pull requests = $1.65 / pull request
Cubic v2 $40 ÷ 9.1 pull requests = $4.40 / pull request
diffox $0.41 / review (the 50 reviews’ charge ÷ 50, addition included)Competitors’ prices
| Tool | Plan | Price | Source |
|---|---|---|---|
| Qodo v2 | Pro Team credit pack | $0.012 per credit; 2,500 credits ($30) ≈ 18 reviews, 5,000 ($60) ≈ 36, 20,000 ($240) ≈ 144: ≈ $1.67 per review. Smaller pull requests use fewer credits. | page |
| Qodo Extended | Pro Team credit pack | Not for sale. In its post of March 2026 Qodo describes Extended as “currently only research results and not yet available in production” and as “an orchestrated multi-agent layer added on top of the default agents”; it does not appear on the pricing page or in the docs. Since it cannot cost less than the standard review, Qodo’s average per review is shown as a lower bound. | page |
| Cubic v2 | Team | $40 per developer per month; 40,000 reviewed lines per developer included. | page |
| Greptile v4 | Pro | $30 per seat per month; includes 50 credits (a base review is 1 credit), extra credits are $1 each. | page |
| CodeRabbit | Essentials | $30 per developer per month; unlimited reviews, with an hourly rate limit. | page |
| Sourcery | Pro | $15 per developer per month. | page |
| Gitar | Core | $25 per user per month; unlimited iterations per pull request. The page is now on Sonar’s site. | page |
| CodeAnt | Premium | $30 per user per month; unlimited pull request reviews. | page |
| Baz | Pro | $30 per active developer per month; unlimited code review agents. | page |
| Augment | Code Review (credit era) | Its own launch post gives an average of 2,400 credits per review ≈ $1.50. The current page bills tokens plus a service fee and gives no per-review figure; the GitHub bot is being retired. | page |
| Cursor Bugbot | Bugbot (usage-based) | By Cursor’s own account an average run costs $1.00–1.50 (May 2026). The board’s run was made in January 2026, under the old $40-per-developer plan. | page |
How we calculated
diffox. In the run of 7 October 2026, our billing records charged $19.93 for diffox’s 50 reviews and $81.52 for diffox ultra’s: averages of $0.40 and $1.63. The $0.40–0.41 and $1.63–1.64 in this post add a small amount per setting. That amount pays for the finding title and fix suggestion that every production comment carries. The benchmark takes findings raw, so those steps did not run in that run; we measured what they were charged over the last 30 days of production records and added it. It raises the price by less than 3%. The price of a review depends on the size of the change reviewed; these figures are the average over these 50 pull requests.
Competitors. We read each competitor’s own public page on 9 October 2026. For tools with a plan price we took the monthly billing price and did not apply annual discounts; for Qodo, we combined the credit price ($0.012) with Qodo’s own estimate (2,500 credits ≈ 18 reviews); for Augment and Cursor Bugbot we used the average the vendor publishes. Qodo Extended is not for sale at the moment: Qodo describes it as a multi-agent research layer, not yet in production, added on top of the default agents, and gives no price; so it cannot cost less than a standard review. Rather than assume a percentage, we show Qodo’s average as a lower bound; the real price may be higher, and every claim we make holds at the lower bound too. We took list prices as published.
Re-reviews. diffox reviews every push to a pull request automatically and reads only the new change; after the first 5 pushes it reviews when you ask (you can change this number). The calculations assume one review per pull request. If your pull requests get two reviews on average, halve the break-even points; that is the most it can change, because the second review reads only the new change. Subscription tools include re-reviews in the monthly fee.
Limits. Each tool’s quality comes from a single run; differences of a few points should not be read as a ranking, which is why we say “the same score”. “First” and “top five” come from the score order of a single run. “Cheapest” rests on the price gap and does not depend on that order: no tool that scores higher than diffox costs less, and the nearest costs about 3.7 times as much. The board’s runs were made on different dates, while we read the prices on 9 October 2026; a vendor may have changed its plan in the meantime.
All the numbers together
| Tool | F2 | Precision, % | Recall, % | Price per review |
|---|---|---|---|---|
| diffox ultra · High sensitivity | 66.5 | 45.2 | 75.3 | $1.64 |
| Qodo Extended* | 65.1 | 67.1 | 64.6 | $1.67+ |
| Cubic v2 | 64.9 | 61.5 | 65.8 | $3.51–5.63 |
| Augment | 64.0 | 59.5 | 65.2 | $1.50 |
| diffox ultra · Medium sensitivity1 | 62.6 | 65.7 | 61.9 | $1.64 |
| diffox · High sensitivity | 60.9 | 44.5 | 67.1 | $0.41 |
| Qodo v2 | 60.6 | 55.4 | 62.0 | $1.67 |
| diffox · Medium sensitivity1 | 57.4 | 62.0 | 56.3 | $0.40 |
| Gitar | 53.6 | 35.5 | 61.4 | $2.19–3.52 |
| Sourcery | 51.5 | 41.0 | 55.1 | $1.32–2.11 |
| Greptile v4 | 51.2 | 50.9 | 51.3 | $2.63–4.23 |
| CodeRabbit | 51.1 | 32.6 | 59.5 | $2.63–4.23 |
| diffox ultra · Low sensitivity1 | 48.8 | 78.4 | 44.6 | $1.63 |
| Cursor Bugbot | 48.6 | 56.9 | 46.8 | $1.00–1.50 |
| diffox · Low sensitivity1 | 41.5 | 75.6 | 37.3 | $0.40 |
| CodeAnt | 40.9 | 40.1 | 41.1 | $2.63–4.23 |
| Baz | 40.5 | 61.5 | 37.3 | $2.63–4.23 |
1 Medium and Low sensitivity are our internal measurement on the same 50 pull requests and 158 bugs: the average of 5 separate runs for diffox ultra and 2 for diffox, rounded. The High sensitivity rows are from the public run. * Qodo Extended is not for sale and has no price: Qodo describes it as “currently only research results and not yet available in production”. We show Qodo’s average per review instead, as a lower bound (+).
The 10 tools whose prices we did not chart
We left 10 of the board’s 21 competitors out of the price charts, because what their vendors publish cannot be turned into a single price per review: it is a wide range, belongs to another product, contradicts itself or does not exist. Rather than fill the gap with a guess, we give what each vendor publishes as it stands. Some of those figures are below diffox’s: GitLab is $0.25 per review, and GitHub Copilot’s estimate starts at $0.05; both score below diffox (52.7 and 58.0). None of these 10 tools scores higher than diffox.
| Tool | F2 | What the vendor publishes | Why it is not charted |
|---|---|---|---|
| GitHub Copilot | 58.0 | An estimated $0.05–1 of AI credits per review at Lite effort, $0.25–5 at Balanced (the default since 28 September 2026). 1 AI credit = $0.01; Actions minutes on private repositories are extra. | A range twenty times wide that moves with pull request size is an estimate, not a price. |
| Gemini | 52.8 | The enterprise version is free during its preview (a Google Cloud billing account is required). | The consumer GitHub app, which the board may have run, was shut down on 17 July 2026. |
| GitLab Duo Code Review | 52.7 | Code Review Flow costs $0.25 per review (0.25 GitLab credits), flat. | The board’s entry carries the name of GitLab’s separate, non-agentic feature, and Martian does not say which one ran. |
| Macroscope | 50.8 | $0.05 per KB reviewed at the default setting, 10 KB minimum: $0.50 for most reviews. | The price depends on the size of the change, and the board’s run (March 2026) came before this billing began. |
| Claude Code (CLI) | 47.7 | Anthropic’s managed Code Review averages $15–25 per review. | Martian does not say which product its two Claude entries ran; one of them is dated before the managed product launched. |
| Devin | 44.8 | Review size per pull request in ACUs: XS at most 2.25, XL above 18. No dollar price. | What an ACU costs is set in each customer’s contract. |
| Kodus | 41.5 | $10 per developer per month ($8 billed annually) plus the token fees you pay your own LLM provider. | The token bill depends on the model you choose; a per-review figure would be a guess. |
| Claude Code | 38.9 | Same as the row above. | Same as the row above. |
| KG | 18.4 | Kilo’s own pages say all of these: free, $0.02–0.15 per review, $0.33 an hour of compute plus model fees, and $0.73–1.34 per review. | The comments of the board’s entry carry Kilo’s signature, but Kilo’s pages give different prices for a review. |
| Graphite | 9.3 | AI reviews are “limited” on Starter and unlimited on Team; no number is given. | No limit is published. |
Frequently asked questions
How do you know diffox is the “cheapest of the top five” and that “every higher-scoring tool costs more”?
Ranked by each tool’s F2 score in its single run, the top five are diffox ultra (66.5), Qodo Extended (65.1), Cubic v2 (64.9), Augment (64.0) and diffox (60.9). The cheapest of the five is diffox at $0.41 per review; the next cheapest, Augment, has a published average of $1.50, about 3.7 times diffox’s. Every tool that scores higher than diffox costs more than diffox; the remaining 18 competitors score below diffox, so none of them can be both cheaper and higher-scoring. No tool scores above diffox ultra. Qodo Extended is not for sale, but it is on the board, so it counts in the top five; leave it out and Qodo v2 (60.6, $1.67) takes its place, with diffox still the cheapest of the five. “First” and “top five” are the score order of a single run; “cheapest” rests on the price gap and is independent of the score order.
What does a review cost with diffox?
The price depends on the size of the change under review and is written in the last line of each review’s summary comment. The $0.41 and $1.64 in this post are the average over the benchmark’s 50 pull requests; on your own pull requests it can differ. Current prices are on the pricing page.
Does the number of findings posted change the price?
Hardly at all. Every posted finding gets a title and a suggested fix, so the price moves only slightly. On these 50 pull requests, the price difference between the three sensitivity settings was at most 1.5%.
How does the bill change if every push is reviewed again?
diffox reviews every push to a pull request automatically but reads only the new change; after the first 5 pushes it reviews when you ask, and you can change that number in the console. The calculations in this post assume one review per pull request. If your pull requests get two reviews on average, the break-even points with a subscription are halved; subscription tools include re-reviews in the monthly fee.
Aren’t competitors cheaper on annual billing?
For some, yes: Sourcery, Gitar, CodeRabbit and CodeAnt offer 20% off on annual billing, Cubic 25%. We compared monthly plans and did not apply annual discounts.
Can I check these numbers myself?
Quality, yes: the result files and commands are at github.com/diffox-bench/code-review-benchmark. Competitors’ prices too: each row’s source is linked in the table. You see diffox’s price on your own pull requests: the last line of each review’s summary comment shows that review’s charge.
The architecture, explained
The third post describes what produces these results: how diffox reviews a pull request.