Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableGitLab Blog→ original

GitLab Duo reviews AI agents' code for $0.25, versus $15–25 at Anthropic

GitLab proposes having its Duo assistant check code written by AI agents such as Claude Code, Codex and Cursor. Duo reviews merge requests, the sets of changes waiting to be merged into the main branch. One review costs $0.25, so 1,000 reviews come to $250. By GitLab's estimate, Anthropic's Code Review runs $15–25 per review, or $15,000 or more for the same volume. That makes it practical to review even small changes, not just large ones.

AI-processed from GitLab Blog; edited by Hamidun News
GitLab Duo reviews AI agents' code for $0.25, versus $15–25 at Anthropic
Source: GitLab Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceGitLab Blog ↗
Editor-in-Chief
Key metric #1Failing CI. Duo fixes broken pipelines on its own, opening a separate MR with the fix.
Key metric #2Security. On the Ultimate tier, Duo suggests fixes for vulnerabilities about 10 minutes after a scan.
Key metric #3Approvals. An approval from Duo does not count toward required approvals.

GitLab proposes a simple idea: let an AI agent write the code, and let the platform handle checking it. The subject is the GitLab Duo assistant, which reviews merge requests (MR), the packages of changes waiting to be merged into the project's main branch. A single such review costs a flat $0.25 at list price, so 1,000 reviews come to $250. For comparison, GitLab cites Anthropic's average Code Review price: $15–25 per review, or $15,000 or more for the same 1,000 reviews. GitLab treats this price gap as an argument for reviewing every change rather than only a sample.

What problem the combination solves

Most teams working with AI agents end up in the same situation. One engineer asks Claude Code to check changes and watch CI (automated builds and tests after each commit), another does everything by hand, and a third never looks at security scanner notifications. As a result, changes in one project move at different speeds and under different levels of control, depending on each person's settings and free time. GitLab proposes solving this not by choosing a "correct" agent, but by moving the routine work into the platform itself.

The company has put together a recommended configuration for Claude Code, Codex, Cursor, Devin and OpenCode. The agent still writes code in the user's editor. Everything from the merge request through to deployment is handled by GitLab Duo.

Setup works through a "skill" for the agent. You paste a short instruction with the repository address into your agent, the agent installs the gitlab-duo-setup skill, connects to GitLab, and enables automatic review and repair of failing pipelines. Before making changes, it shows a plan, and nothing is applied without your consent. A group owner (the Owner role) can configure the whole group, and a maintainer can configure a single project. An additional MR follow-up skill brings review comments and CI results back into your session when you reopen the editor. Work continues in GitLab even after you have closed the editor.

What the team gets

The price of a review is the main argument, but not the only one. The comparison with Anthropic was made by GitLab itself, and terms may differ between products, but the order of magnitude of the difference is clear. At $0.25 per review, even routine changes can be sent for review.

Other scenarios look like this:

  • Failing CI. When a pipeline turns red, the Fix CI/CD Pipeline action starts automatically. Duo proposes a fix, opens a separate MR with the correction, or explains what it found if there is no reliable fix. The fix still has to pass CI.
  • Security. On the Ultimate tier, Duo flags likely false positives from static application security testing (SAST) and secret detection. For suitable SAST findings, it opens an MR with a fix, usually about 10 minutes after the scan.
  • Tasks. An agent can hand an assigned issue to the Developer flow: it reads the task and the repository, writes the change, and opens an MR on your behalf. There is no need to copy context into a prompt.

Control and limitations

Review rules are set once in GitLab and apply equally to code from an agent and from a human. Approval policies can require human confirmation for every MR or add approvers when a scan finds a vulnerability. Pipeline execution policies guarantee that mandatory jobs run. Tool guardrails make the agent ask a human before a sensitive action.

Agent permissions are constrained by a composite identity mechanism: each action is tied to both the agent and the person who launched it, and access is limited to what both are allowed to do. Calls through the GitLab MCP server can be viewed in the audit log if retention is enabled. The model for each scenario is chosen from options by Anthropic, OpenAI, Google and open-weight providers.

There are caveats too. An approval from Duo does not count toward required approvals, so the final decision stays with people. Some capabilities, including vulnerability analysis, are available only on Ultimate. The cost and fix-time figures come from the vendor itself, and the material contains no independent measurements of quality. How many problems Duo finds turn out to be significant will have to be verified on your own repositories.

What to do about it

If a team already works in GitLab and uses different agents, it makes sense to try the combination on one project. Run reviews on a stream of real MRs, compare the findings with what your people catch, and see how many failing pipelines Duo fixes without intervention. If the quality is acceptable, you get a single review standard for everyone, regardless of whose agent and whose settings the change's author is using.

Story

Key Questions Answered

What problem the combination solves?

Most teams working with AI agents end up in the same situation. One engineer asks Claude Code to check changes and watch CI (automated builds and tests after each commit), another does everything by hand, and a third never looks at security scanner notifications. As a result, changes in one project move at different speeds and under different levels of control, depending on each person's settings and free time.

What the team gets?

The price of a review is the main argument, but not the only one. The comparison with Anthropic was made by GitLab itself, and terms may differ between products, but the order of magnitude of the difference is clear. At $0.25 per review, even routine changes can be sent for review. Other scenarios look like this: Failing CI. When a pipeline turns red, the Fix CI/CD Pipeline action starts automatically.

What are the key details on Control and limitations?

Review rules are set once in GitLab and apply equally to code from an agent and from a human. Approval policies can require human confirmation for every MR or add approvers when a scan finds a vulnerability. Pipeline execution policies guarantee that mandatory jobs run. Tool guardrails make the agent ask a human before a sensitive action.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…