Touting lower costs and faster speed
Cognition, which makes the AI programmer "Devin," on Nov. 5 (Nov. 6 Japan time) announced "Devin Review 2.0," a new version of "Devin Review," in which AI checks code changes. What it checks are pull requests (PRs), which developers submit before merging changes into the main codebase. The company said on X that it had become up to 5 times cheaper and 1.5 times faster. It describes the lower cost as bringing the tool to a level where it can be used on every PR.
Introducing Devin Review 2.0: the best way to review PRs with Devin. Now up to 5× cheaper, 1.5× faster, and with new features and integrations.
The company's concern is that, as the number of PRs grows, code review has become a new bottleneck in development. Devin Review is a product released on Jan. 21 that organizes changes and detects bugs.
"5 times cheaper" compares with its own product in August
According to Cognition's blog, the cost of reviewing a single PR has fallen 82% since August. The blog's headline says it is 5x cheaper since August, and the "up to 5x" in the X post appears to refer to the same comparison. The baseline for the comparison is the company's own product as of August, and it did not give the amount at that time. It also did not give figures comparing it with other companies' products. Whether the 82% is an average could not be confirmed from the blog either.
As for current costs, the company says a single review generally costs $0.25 to $0.75. Even including all pushes to the same PR, about 80% of PRs come in under $1, it says. Averaged across all customers, the cost per bug that led to a fix is about $2.25, it said. It attributed the lower costs to revisiting the combination of models used and optimizing the surrounding systems that run the AI.
On speed, the median time per review fell from 139 seconds to 94 seconds, the company said. Even after cutting costs, the average number of bugs found per PR rose from 1.22 to 1.32, the company said.
No third-party verification found, and no false-positive rate given
All of the figures above come from Cognition, and as of Nov. 6 no third-party verification of them could be confirmed. The blog gives no detailed explanation of which customers or how many cases were included in the tally, and it names no customers. AlphaSignal, which reported the announcement, also noted that the figures were provided by the company and that results may vary by repository and the content of changes.
The number of detections rose, but no figures showing the quality of the comments were released. False-positive rates, rates of missed bugs and the share of comments accepted by developers could not be confirmed from the blog. For companies adopting the tool, the accuracy of comments on their own code will be the focus.
Re-reviews cover only changed parts
When the same PR is reviewed again, only the parts that changed are examined. This incremental review applies to all PRs by default. The company said that narrowing new comments to the latest changes also reduces unnecessary comments.
Reviews are started by commenting "/devin review" on a GitHub or GitLab PR, or from the Devin Cloud and Devin Desktop interfaces. Users can also create rules to run reviews automatically, choosing to run them on every push or to narrow the scope by file location, branch or label. A spending cap per PR can also be set to keep costs from ballooning.
Fixes for comments handed to Devin, with conditions
The blog cited 2 new features. "Ultra" is a setting for billing and security-related code that reviews more carefully at a higher cost. It did not say how much more it costs. "Custom Reviewers" lets teams create their own checks using instructions they write; it is offered as a beta and requires beta access.
For comments it finds, Devin investigates and decides whether to fix, dismiss or send them back to a person. According to the company, reviews can run on PRs created by any route but work best when combined with Devin. The official documentation, however, says the automatic "Auto-Fix" requires the Devin session that created the PR.
Many of the features had been rolled out gradually before 2.0. The changelog lists effort settings on Oct. 5 and custom reviewer permissions on Oct. 14. The specific partners for the "integrations" cited in the X post could not be confirmed from the blog. Devin Review's supported platforms include Bitbucket Cloud, which was added in October.
- Devin Review released; free during early release periodOfficial
- Autofix, in which Devin fixes review comments, announcedOfficial
- Baseline for the 82% reduction; amount at the time undisclosedOfficial
- Review effort settings in changelogOfficial
- Custom reviewer permissions in changelogOfficial
- Starting reviews from GitLab comments supportedOfficial
- Bitbucket Cloud supportedOfficial
- Devin Review 2.0 announcedOfficial
Pricing structure and availability unconfirmed
The new blog post does not say which plans include Devin Review. Devin's pricing page lists Free at $0, Pro at $20 a month, Max at $200 a month, Teams at $80 a month plus $40 per seat, and Enterprise with pricing on request, but it does not mention Devin Review. Whether $0.25 to $0.75 is the amount charged to users on each plan could not be confirmed either.
According to the official documentation, reviews consume "ACUs," Devin's usage units, from a shared enterprise pool. Organizational management and consumption controls require an Enterprise contract. However, it does not show how ACUs correspond to $0.25 to $0.75.
At its January launch, the company said the tool was free during the early release period, and in a March post on X it also described it as completely free. Whether any free usage remains could not be confirmed from the latest materials. Supported platforms are GitHub, GitLab, Azure DevOps, Bitbucket and Perforce Swarm, and available features vary by platform.
Competitors charge by different units
AI code review tools are billed in different units by each company, so they cannot be compared with Cognition's figures under the same conditions. Anthropic's "Claude Code Review" has no flat fee and charges based on the number of tokens used. GitHub Copilot's code review incurs costs in 2 ways: AI credits and GitHub Actions run time. CodeRabbit charges a monthly fee per developer, with its cheapest plan, Essentials, at $24 billed annually or $30 billed monthly.
Between pricing per developer and pricing per PR, which is cheaper depends on the number of PRs. Claude Code Review's documentation also says that running it on every push increases costs in proportion to the number of pushes. Cognition's incremental reviews and spending caps are likewise mechanisms to hold down costs when re-reviews pile up.
Cognition piles up announcements
Cognition revamped "Devin Security," which handles vulnerabilities, on Oct. 26, and the next day published "FrontierCode 2.0," an evaluation of AI coding tools. On Nov. 4, it released "SWE-2 Flash," which it says is priced at one-fifth of higher-end models. It gave no accuracy or cost figures for Devin Security, and no third-party verification of SWE-2 Flash's performance or pricing has been confirmed.
The company raised more than $2 billion in September at a valuation of $48 billion, and on Sept. 25 announced that its annualized revenue had topped $1 billion. Revenue and user numbers for Devin Review alone could not be confirmed in public materials.
