GPT-6 Astra catches 22% more bugs than Opus 5, at 2.5x the cost
CodeRabbit's eval: Astra's edge jumps to 33% on cross-file bugs. A review runs $1.50 vs Sol's $0.60 — but batch halves it, and zero-retention is on offer.
Copy markdown
22% more bugs — 33% on cross-file
In CodeRabbit's eval, Astra flagged 61.3% of actionable bugs versus 50.2% for Opus 5 and 59.0% for GPT-5.6 Sol. On harder cross-file reviews the lead widens to 57.1% — a third more than Opus 5.
What a review actually costs
At $10/$50 per million tokens, a 100K-in/10K-out review runs about $1.50 — 2.5x GPT-5.6 Sol's $0.60, and 47x Luna's $0.032. Opus 5 and Fable 5.1 match Astra's $1.50.
Route it, don't default to it
The economics favor task-routing: send cross-file and distributed-reasoning reviews to Astra, keep routine self-contained diffs on cheaper models. Optimize cost-per-caught-bug, not per-token.
Batch cuts the bill in half
Astra's Batch API tier runs $5/$25 per million — 50% off the standard rate. For overnight PR sweeps that don't need instant turnaround, that turns a $1.50 review into $0.75.
The retention fine print
Astra offers zero data retention for eligible API customers; Anthropic's Fable defaults to 30-day retention with a ZDR option. Worth checking before you pipe proprietary code through review.