GPT-6 Astra Records 4% Higher Actionable Bug Coverage Than GPT-5.6 Sol

CodeRabbit evaluated GPT-6 Astra on labeled bug detection across standard and cross-file pull requests. Astra reached higher coverage on distributed evidence tasks, showing 20% gains over Sol and 33% over Opus 5 in the harder subset. The test measured findings developers could act on directly rather than raw recall.

Pricing data from OpenAI's September 2026 API sheet lists Astra at $10 per million input tokens and $50 per million output tokens. At 100k input and 10k output tokens this produces 2.5 times the cost of Sol and 4.7 times Terra. Larger context alone does not explain the cross-file lift; reasoning trace quality appears decisive.

Teams running repeated reviews on large codebases face a direct trade-off between per-task accuracy and token budget. Astra's edge concentrates where changes touch multiple modules, matching OpenAI's stated positioning for multistep agent workflows. Routine single-file reviews show smaller separation, limiting blanket replacement of cheaper models.

Deployment records from similar model transitions indicate that sustained usage requires measurable reduction in total tokens per merged PR before premium pricing scales. No public benchmark yet isolates whether Astra's context handling reduces downstream defect rates in production systems.