AI21 Labs says it cut GitLab Duo Code Review’s measured cost from $0.589 to $0.0135 per pull request—a 97.7% drop—by running all four stages on DeepSeek V4 Flash and compacting the prompts. On Martian’s Code Review Bench, the final setup scored 47.2% F1, compared with 46.4% for the Sonnet 5 baseline.
The test used 50 public pull requests, not GitLab production traffic or customer data. AI21 says three requests exceeded its DeepSeek deployment’s 256K-token limit without compaction and fell back to Sonnet; compaction brought every review within the limit. The settings were tuned and scored on the same requests, so performance on other codebases is untested.
Why it matters
Teams running code-review agents can treat the results as evidence that model choice and prompt size may affect per-review spending, but AI21’s test does not establish the same savings or quality on other codebases or with GitLab’s newer model.
