AI Code Review Tools Race Ahead as Developer Distrust Hits 46%

The new bottleneck is checking AI-written code

For most of the past two years, AI coding companies competed on how much code their tools could write. By October 2026, they are competing on how well that code gets checked. Anthropic, GitHub and Cursor all have pull-request review products on the market now. In the editor itself, reporting describes Zed, Cursor, VS Code and Claude Code adding or expanding review controls that let developers accept or reject AI edits hunk by hunk instead of file by file. That shift reportedly landed within days of a Claude Code update on October 17.24

The reason behind all of this shows up in one survey number. Stack Overflow's 2025 Developer Survey found that 46% of developers distrust the accuracy of AI tool output, up from 31% a year earlier.26 Over the same period, the share using or planning to use AI tools rose from 76% to 84%.26 So developers are using these tools more and trusting them less, and the vendors' answer is more automated checking. The main open question is whether AI reviewing AI can close that gap.

What the trust gap actually measures

The figures in the survey's headline get mixed up a lot, so they are worth separating. The 46% is distrust in accuracy specifically. Only 33% said they trust accuracy, and roughly 3% said they trust it highly.25 A different question about overall trust in AI tools came in at 29%, down from about 40% in 2024.2125 Different outlets lead with different numbers, but they agree on the direction: confidence is falling while adoption keeps climbing.

The survey also shows why the gap matters for review. Two-thirds of respondents named "almost right, but not quite" answers as a problem, and 45% said debugging AI-generated code takes a long time.2327 Experienced developers were the most skeptical. Only about 2.6% of them highly trust AI output, while around 20% highly distrust it.2329 Distrust, in other words, does not come mainly from people who don't know the tools. It comes from the people who spend the most time fixing what the tools get wrong.

Agents make the problem bigger. Only 31% of developers were using AI agents at the time of the survey, and 38% said they had no plans to.24 Agents are the part of the stack that produces large, multi-file changes that nobody has read yet, so they create the most review work.

Anthropic: expensive, thorough, and generally available

Anthropic's Code Review for Claude Code became generally available to Team and Enterprise organizations on October 15.32 It came out of a research preview first announced on March 9.33 The system runs several agents in parallel over a diff and the code around it. A verification step then filters out false positives, and the remaining findings are deduplicated, ranked by severity, and posted to the pull request.31 According to Anthropic, most reviews now take 6 to 10 minutes, compared with about 20 minutes during the preview.32

Anthropic's internal numbers are what it uses to sell the product. Before deployment, 16% of the company's pull requests received substantive review comments; afterward, 54% did. Engineers marked fewer than 1% of findings as incorrect.36 All of those figures come from Anthropic's own codebase and haven't been checked independently, so they are best treated as an upper bound until outside teams report their own results.

The product's limits are deliberate. A review never approves or blocks a pull request, and its check run never reports a failure, so humans keep the merge decision.31 It is also expensive. During the preview, reviews were billed on token usage and averaged $15 to $25 each.35 To get teams to try it, Anthropic is offering eligible Enterprise customers $10,000 to $50,000 in credits through January 15, 2027.38 There are also restrictions: the service works only with GitHub, and it is unavailable to organizations with Zero Data Retention or HIPAA configurations.31

This background also complicates some popular comparison guides. As recently as March, one of them listed Claude Code as having "no native review."48 That is no longer true, and buyers relying on older comparison tables should keep in mind how quickly this category is changing.

GitHub: review as a feature of the platform

GitHub took a different approach. On March 5, Copilot code review moved to an agentic, tool-calling architecture. It now pulls in broader repository context and is generally available on the Pro, Pro+, Business and Enterprise tiers.41 Since then GitHub has added several features. In June it began reading AGENTS.md files.43 In July it brought agent skills and read-only MCP connections to general availability.42 In August it added coverage for very large pull requests and for pull requests opened by bots, including GitHub's own Copilot cloud agent. Developers can also now tag a dismissed comment as "Incorrect" or "Won't fix."49

That last change matters. A reviewer that checks code written by its own sibling agent turns the loop between writing and reviewing into a single product. GitHub's changelog also lists an entry saying Copilot code review can now approve pull requests.41 On this point GitHub and Anthropic have clearly split: Anthropic keeps merge authority with humans, while GitHub is moving toward letting its tool sign off. My reading is that GitHub is betting that tight integration with the platform will make automated approval acceptable before trust data supports it.

Coverage of GitHub's numbers is uneven. Third-party write-ups say Copilot has done more than 60 million reviews, with 71% producing actionable feedback and an average of 5.1 comments each.47 Some guides say the feature requires Business or Enterprise, which contradicts GitHub's own changelog listing the individual Pro tiers.4441 On pricing, multiple outlets report that code review began consuming AI credits and Actions minutes on June 1, so it is no longer effectively free inside a seat.1344

Cursor: from a $40 seat to paying per run

Cursor's Bugbot followed the same pricing path. It used to cost $40 per user per month. Cursor announced on May 11 that it was moving to usage billing, and existing subscribers switch at their first renewal after June 8.16 Cursor now says the average review costs $1.00 to $1.50, depending on pull-request size and complexity.11 There is a catch: by default Bugbot reviews every update to a pull request, so a branch that gets pushed several times is billed several times.1314 One analysis concluded the new model is cheaper for teams that open few pull requests per developer and much more expensive for teams that open many.12 Separately, one pricing tracker reports that SpaceX acquired Cursor in August and that Bugbot is still on sale.13

Market share is shifting away from the editor

Usage data helps explain why review has become the place where vendors compete. A mid-2026 report found that 39% of professional developers used Claude Code at work, up from 18% in January. Copilot was at 21% and Cursor at 12%, and both were declining.3 Comparison guides largely agree on how the tools differ: Claude Code for deep multi-file reasoning, Copilot for GitHub-native teams, and Cursor for interactive editing.210 As agents write more code outside the editor, owning the step where changes are checked may be what keeps customers loyal.

What it means

Pricing models are converging on per-review costs that can be predicted and capped. Products are converging on repository-level instruction files such as REVIEW.md, AGENTS.md and SKILL.md. Even so, I don't think the trust gap will close just because review tools ship. Anthropic's own documentation warns that a review with no findings does not mean the code is error-free or secure.31 Developers say their biggest frustration is code that looks correct and isn't, and a second AI that also produces plausible-sounding answers doesn't fix that.

What could actually rebuild confidence is measurement. GitHub's new "Incorrect" resolution tag and Anthropic's analytics dashboards can start producing real false-positive and miss rates at scale, beyond vendor anecdotes.4936 Until outside teams publish that kind of data, the sensible approach for engineering leaders is the one Anthropic built into its product: use AI review as a first pass, and leave the merge decision with a human.