A change can look safe in one file and break another. CodeRabbit tested GPT-6 Astra on that kind of code review and reports its largest gains on difficult cross-file cases.
In CodeRabbit's evaluation, Astra produced actionable findings for roughly 4% more labeled bugs than GPT-5.6 Sol and 22% more than Claude Opus 5 overall. On the harder cross-file subset, the reported relative improvements were 20% and 33%, respectively. These are CodeRabbit's early internal measurements, not an independent ranking of all models or a prediction of how many bugs a team will prevent.
Astra is also more expensive at published standard API rates. CodeRabbit lists $10 per million input tokens and $50 per million output tokens for Astra, against $4 and $20 for Sol at the prices checked on September 4. Its example using 100,000 uncached input tokens and 10,000 output tokens costs $1.50 on Astra versus $0.60 on Sol. Caching, retries and reasoning-token use can change the total for a real task.