Cognition says Devin Review is now the default pull-request review tool when teams work with Devin, and that a 2.0 refresh has cut the cost of reviewing a PR by 82% since August. The 5 November 2026 product post puts a typical review at $0.25 to $0.75, with about 80% of PRs under $1 including all pushes. Devin’s Review docs describe the same product surface: incremental reviews, effort modes, spend limits and Custom Reviewers in beta.
Those dollar and percentage figures are Cognition’s own customer averages. They are not an external lab score, and they are not a claim that Devin Review beat named rivals in a shared public harness.
What Cognition says changed in 2.0
On Cognition’s numbers, bugs found per PR rose from 1.22 to 1.32, median latency fell from 139 seconds to 94 seconds (about 1.5× faster), and the average cost attributed to each fixed bug across customers is about $2.25. Reviews of changed parts run incrementally by default after pushes. Teams can pick Devin Review Ultra effort, set a per-PR spend limit, and use auto-review rules such as a GitHub or GitLab /devin review comment. Custom Reviewers remain in beta. Docs also note multi-provider support and Normal versus Ultra effort settings.
None of that package is SWE-2 Flash. Cognition’s separate SWE-2 Flash model post is a different product line; this article covers only Devin Review’s PR-review path.
How this sits next to other TSN coding-agent posts
TSN has already tracked Cognition’s Devin memory and “dreaming” features as an earlier Devin product step. Rival agent marketing shows the same pattern of vendor-owned scores: Factory says Droid 1.0 beats Codex and Claude Code in its own tests. Cursor’s lane is different again — Live Apps let cloud agents build and host web apps — so it is adjacent tooling, not a Review substitute.
Reading Devin Review 2.0 as an industry-wide “cheapest PR review” crown would confuse Cognition’s internal trend line with a cross-vendor benchmark. The useful comparison on TSN’s shelf is process: more coding agents now sell review as a metered workflow with spend caps, not only as chat autocomplete.
Company averages are not a shared leaderboard
For engineering leads, the 82% cost cut and $0.25–$0.75 typical band answer a budget question inside Cognition’s installed base since August. They leave open how those PRs were sampled, which languages dominated, and whether Ultra effort was the default. The jump from 1.22 to 1.32 bugs per PR is a small absolute move; Cognition presents it as an improvement, not as proof that reviewers can skip human sign-off.
Spend limits and incremental reviews are the operational hinge. If most of the dollar savings come from reviewing only the delta on later pushes, that is a product design choice as much as a model quality story. Custom Reviewers in beta hint that teams still want organisation-specific rules on top of the generic pass — another sign that “default tool” does not mean “finished policy.”
UK readers get no UK pricing page in these sources. The blog and docs speak in US dollars and GitHub/GitLab workflows; any UK or EU data-residency story would need a separate Cognition disclosure.
What to watch next
Three public breadcrumbs would harden the file: Cognition publishing methodology for the August-to-now cost and latency series; Custom Reviewers leaving beta with named enterprise controls; or a third-party evaluation that retests Devin Review against other PR bots on a fixed corpus. Until then, Devin Review 2.0 is a Cognition-reported cost, speed and bug-count refresh for its default Devin PR reviewer — not an independent ranking.
Sources
Primary:
Related TSN:
- Devin Gets a Memory That “Dreams”, and Hark Pro Uses Your Computer for You
- Factory Says Droid 1.0 Beats Codex and Claude Code in Its Own Tests
- Cursor Live Apps: Cloud Agents Build and Host Web Apps
Related stories
- Devin Gets a Memory That “Dreams”, and Hark Pro Uses Your Computer for You
- Factory Says Droid 1.0 Beats Codex and Claude Code in Its Own Tests
- Cursor Live Apps: Cloud Agents Build and Host Web Apps
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
