CodeRabbit measured GPT-6 Astra catching 33 percent more cross-file bugs than Opus 5, at 2.5 times the cost of Sol

CodeRabbit published an evaluation of OpenAI's GPT-6 Astra for code review on 4 September, written by Erik Thorelli and Erfan Al-Hossami. The finding worth reading is the gap between the overall number and the hard-case number, because they point in different directions on whether to pay for it.

CodeRabbit measures what it calls actionable bug coverage: how many labelled bugs a model catches through findings a developer can act on. On that measure it reports Astra catching about 4 percent more than GPT-5.6 Sol and 22 percent more than Opus 5. On the harder cross-file subset, where a change looks correct in isolation but breaks something elsewhere, it reports Astra's advantage growing to 20 percent over Sol and 33 percent over Opus 5.

These are CodeRabbit's own results on its own labelled set, and the company sells code review, so it has a stake in the conclusion. The post does not state a sample size. CodeRabbit itself calls the result early and directional, and says explicitly that it does not establish an overall ranking of review quality or predict any team's defect rate.

The price gap is larger than the quality gap

The pricing section is the part a team can act on, because it uses published rates rather than CodeRabbit's own measurements. Astra's standard API rates are $10 per million input tokens and $50 per million output tokens. CodeRabbit builds an illustrative task at 100,000 uncached input tokens and 10,000 billable output tokens, which gives $1.50 for Astra, $0.60 for Sol, $0.32 for Terra and $0.032 for Luna. At that fixed usage Astra costs 2.5 times Sol, about 4.7 times Terra and about 47 times Luna. Claude Fable 5.1 carries the same $10 and $50 base rates as Astra, though caching prices differ. The figures were checked on 4 September 2026 and exclude caching, retries, tools and regional uplifts.

CodeRabbit is careful about what that comparison does not settle. Holding token usage constant is what makes the rates comparable, and real tasks do not hold it constant. A model that finishes in fewer tokens or fewer attempts narrows the gap, and the post notes that OpenAI reports lower estimated task costs for Astra in some of its own evaluations despite the higher token price. That is OpenAI's claim about OpenAI's model, and CodeRabbit passes it on as such.

The shape of the result is the routing rule

Read the two numbers together and they describe a selection rule rather than an upgrade. A 4 percent gain across all reviews at 2.5 times the price is a bad trade. A 20 percent gain on reviews where the evidence sits in several files at the same price may be a good one.

That makes it a per-review decision. The reviews worth spending on are the ones where a change's consequences are distributed: a shared interface, a migration, anything touching code the diff does not show. Routine single-file changes are already handled by a cheaper model, and CodeRabbit's own overall number is the evidence for that.

The honest way to settle it is the one the post recommends: run both models against the same pull requests and compare the cost per caught bug, since that is the figure the token price never gives you. Neither the price list nor a capability score answers the question on its own.

Source

GPT-6 Astra in code review: Gains, privacy, and cost, CodeRabbit, 4 September 2026.