$ git diff main -- methodology.md
The site tracks 19 AI code review tools. Pricing is taken from each vendor's own pricing page and re-verified on the date in the magenta pin. Token math uses Claude Sonnet 4.6 list rates of $3 input and $15 output per 1M tokenssrc. That is a deliberate ceiling: Claude Sonnet 5 lists lower at $2 input and $10 output, so a team on the current generation pays about a third less than the figures here. Quote-only vendors are listed but excluded from $/PR ranking. No accuracy score is invented.
1+Every seat rate is taken from the vendor's own pricing page. Third-party blogs are not used as a primary source.
2+The pricing page URL and the verified date sit in the magenta pin next to the claim, on every page.
3-If a vendor publishes only a price range ("starting at X") we use X and label it "cheapest published" rather than "list".
4-Enterprise discount you negotiated with your rep is not in here. Add a haircut as you see fit.
1Flat per-seat:
monthly = seat_rate * team_size. Per-PR derived as monthly / (PRs_per_wk * 4.33).2Flat whole-team (Augment Code, Git AutoReview):
monthly = flat_rate, with no multiplication by seat count, because that is what the vendor publishes. Both plans have a headcount ceiling (50 seats and 25 members), and above it neither vendor publishes a price, so the calculator says so rather than extrapolating.3Per-PR token meter (Claude Code Review, any BYOK path):
token_cost = LOC * 24 * $3/M + LOC * 2 * $15/M. 24 input tokens per LOC reflects PR diff context-window inflation; 2 output tokens per LOC reflects review-comment density. Both are deliberate over-estimates so the headline does not under-promise.4Bundled (Copilot Code Review): $0 marginal if Copilot Business is already in the budget; full Business seat cost otherwise.
1+Launchable and Diffblue Cover are listed on /scorecard and the calculator with the label "quote-only".
2+Usage-billed products get the same treatment. Cursor BugBot publishes no rate of its own since Cursor retired the seat SKU in May 2026, so it carries no number and sits out of the ranking rather than being modeled from an average run cost.
3-We do NOT publish an inferred rate. No "assumed $50/dev/mo" or "typically $X".
4-Sister sites have been burned inferring rates from leaked deal sizes. We are not doing that.
1OpenSSF Security Baseline and SWE-bench are referenced as the only third-party benchmarks worth pointing atsrcsrc.
2-We do not roll them into a single "accuracy score". Numbers move per release; a snapshot rank misleads.
3-Vendor-published benchmarks (every vendor has them, every vendor wins their own) are quoted on the per-tool page with the vendor labelled as the source. Reader can discount accordingly.
1+Where a vendor confirms figures in writing, the tool page carries a dated note: "Confirmed with [vendor] on [date]", or "Confirmed and corrected directly with [vendor] on [date]" where their reply produced fixes here.
2+The note itemises exactly what the vendor confirmed and covers nothing else. Facts we read from their published pages stay labelled as that, with the pin and the date, because they are a different class of evidence.
3-A vendor correction is not applied on trust. Every claim is re-derived from the vendor's own published pages first, and anything we cannot find there does not go on the page.
4-Suggested wording from a vendor is input, not copy. Verdicts and buyer-fit lines are written here, and a vendor's claims about a competitor carry no weight at all: those are checked against the competitor's own rates.
1Pricing pages get re-verified on a rolling 60-day cadence. The date in each magenta pin is the actual last check, not the page build date.
2/thread is the live log of changes that triggered an out-of-cadence refresh (Claude Code Review launch, Copilot AI Credits, the retirement of the Cursor BugBot seat, Macroscope v3 reprice).
1-No affiliate links. Sister sites carry them; this one does not, because vendor-paid placement would compromise the ranking it exists to publish.
2-No "Editor's Choice" badge. The ranking IS the opinion.
3-No social-proof testimonial blocks. The buyer is a Staff+ engineer who does not believe quote walls.