Entelligence conducted a comparative study between the low-cost GPT-5.6 Luna and the high-performance GPT-6 Astra to evaluate the feasibility of using inexpensive models for automated code reviews. Using 50 public pull requests from the AI-Code-Review-Evals organization, the benchmark evaluated the models on their ability to identify correctness, security, concurrency, resource, and error-handling bugs.
GPT-5.6 Luna vs. GPT-6 Astra: Low-Cost Model Identifies 75% of Bugs at a Fraction of the Cost
The results demonstrated a significant trade-off between cost and precision. Luna cost approximately 3.6% of Astra's price but identified 75% of the verified bugs that Astra caught. While Luna proved efficient for general correctness—finding 39 logic and data bugs compared to Astra's 47—it struggled with specialized code. In Keycloak, an identity and access management server, Luna's verification rate dropped to 50% compared to Astra's 93%, particularly regarding authentication and permission logic.
Regarding security, Luna identified 9 security bugs, compared to Astra's 19. The study concluded that while Luna is highly cost-effective for everyday correctness checks, high-performance models remain necessary for security-sensitive paths where the cost of a missed bug is high.
Sources
- GPT-5.6 Luna vs. GPT-6 Astra: Is a $1.20 Model Good Enough for Code Review? （Hacker News Frontpage, 2026-09-14）
- AI-Code-Review-Evals