CodeRabbit put GPT-6 Astra on code review and priced the gain at $1.50 a review

CodeRabbit published an evaluation of GPT-6 Astra on code review on 4 September, and it did the thing most model comparisons skip: it printed the price next to the score.

What it measured

The metric is actionable bug coverage, meaning the share of real problems in a pull request that the model surfaced as something a developer would act on. CodeRabbit reports GPT-6 Astra at 61.3%, GPT-5.6 Sol at 59.0%, and Opus 5 at 50.2%.

The gap widens on the harder cases. On cross-file reviews, where the bug only shows up when you hold two files in mind at once, CodeRabbit reports Astra at 57.1%, Sol at 47.6% and Opus 5 at 42.9%. That is the class of defect a human reviewer misses most often, and it is where the spread between models is largest.

What it costs

CodeRabbit priced an illustrative task of 100,000 uncached input tokens and 10,000 billable output tokens. GPT-6 Astra came to $1.50, GPT-5.6 Sol to $0.60, and GPT-5.6 Terra to $0.32.

So the top score costs about two and a half times the second and roughly five times the third. On the overall number that buys 2.3 percentage points. On cross-file review it buys 9.5.

Read the caveats, because CodeRabbit wrote them

This is a vendor publishing numbers about models it sells access to, so the attribution matters and so does what the company says about its own limits. CodeRabbit states the results "do not establish an overall ranking of review quality, predict a team's defect rate, or promise the same gain on every pull request", and calls them early and directional. It does not disclose how many pull requests were in the sample or how the ground truth was decided, which is the number that would let anyone else judge the spread.

On data handling, CodeRabbit says neither it nor its model providers train on customer code from private reviews, and that Astra supports zero data retention for eligible API customers.

Where the money goes is now a per-file question

The shape of these numbers argues against picking one review model and leaving it there. If the expensive model's advantage sits mostly in cross-file reasoning, then paying for it on a one-file typo fix is waste, and paying for the cheap one on a change that touches six modules is a false saving.

That is a routing rule, and it needs something to route on: the size of the diff, how many files it crosses, whether it touches code paths that have broken before. A team that already tracks which pull requests produced incidents has the input for that rule sitting in its own history.

The honest limit is that these are one vendor's figures on an undisclosed sample. The useful part is the method: score and price reported together, on the same task, with the caveats attached. Any team can run the same comparison on its own pull requests, and its own defect history is a better ground truth than a benchmark.

Source

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