Starter
Try AI PR review.
- 2 reviewers
- Code Review
- Security Review
- 3 seats included
- 15,000 credits to start, then 5,000/mo
- About 70 reviews to start, then about 25 a month
- Unlimited repos
For GPT 5.6 code review we ran one model. GPT 5.6 Luna Pro is the most precise model we have run on code review, and the one that finds the fewest bugs. On 50 open-source pull requests with 158 known issues, 73% of its findings matched a known issue, but it found only 46: an F1 of 41.6%, for about 12 cents a review in a median 107 seconds.[1]
What lands on your pull request · Code Review
billing/checkout.ts:42 · A timed-out response can charge the order twice- await payments.charge(order);+ await payments.charge(order, {+ idempotencyKey: order.id,+ });
On the code review benchmark it was the only one of our five models to flag a Discourse controller that built SQL by pasting in host names, open to SQL injection.[2]
46 of 158 known issues found, with 17 findings that matched none. Muse Spark 1.3 found 79 and DeepSeek V4.1 Flash 77.[1] Every model's score.
Code ReviewOn Starter, Pro and Max
Every finding it posted on 13 security pull requests matched a known issue, but it found 3 of 18: an F1 of 28.6%, last of the seven models. GLM 5.3 Flash found 10.[3]
Security ReviewOn Starter, Pro and Max
10 of 15 known control gaps found, an F1 of 64.5% against 66.7% for the leaders, DeepSeek V4.1 Flash and GLM 5.3 Flash.[4]
SOC 2 ReviewOn Max
It scored 54.5% on CI/CD, fourth of five models;[5] 50.0% on API performance, fourth of six;[6] 36.0% on logs, last of five;[7] and 35.3% on AI architecture, fourth of eight.[8] We did not run it on docs or cost reviews.
The AssignmentSource dataclass set its queued field's default to timezone.now(), which runs once when the class is defined, so every object shares one time. Luna Pro posted no finding on that pull request. The other four models and all four review apps on that case flagged it.[9]
getSubGroupsCount() returned null although the interface's Javadoc says it never does, which can throw a null pointer error in a caller expecting a count. The benchmark rates it Critical. The other four models caught it, and three of the four review apps missed it as well.[10]
One pagination branch handed a negative cursor offset directly to queryset slicing, and the benchmark rates that Critical too. All four other models caught it, and so did three of the four review apps.[11]
We picked these three because the other models caught them, so they are examples and not a sample. The totals above are the sample, and on code review they say most of what it reports is real and most of what is there goes unreported. That precision does not travel, either, because on three of the seven review types in the table fewer than half its findings matched a known issue. We would run it as an extra pass beside a model that finds more, and would not read silence from it as a clean pull request.
| Review | Model | F1 | Precision | Recall | $ per review | Median time | Rank |
|---|---|---|---|---|---|---|---|
| Code Review50 of 50 PRs | GPT 5.6 Luna Pro | 41.6% | 73.0% | 29.1% | $0.12 | 1m 47s | 5 of 5 |
| Security Review13 of 13 PRs | GPT 5.6 Luna Pro | 28.6% | 100.0% | 16.7% | $0.11 | 1m 16s | 7 of 7 |
| AI Architecture Review11 of 11 PRs | GPT 5.6 Luna Pro | 35.3% | 27.3% | 50.0% | $0.05 | 51 sec | 4 of 8 |
| Log & Errors Review17 of 17 PRs | GPT 5.6 Luna Pro | 36.0% | 47.4% | 29.0% | $0.05 | 44 sec | 5 of 5 |
| API Performance Review18 of 18 PRs | GPT 5.6 Luna Pro | 50.0% | 54.5% | 46.2% | $0.03 | 28 sec | 4 of 6 |
| SOC 2 Review20 of 20 PRs | GPT 5.6 Luna Pro | 64.5% | 62.5% | 66.7% | $0.07 | 37 sec | 3 of 5 |
| CI/CD Performance18 of 18 PRs | GPT 5.6 Luna Pro | 54.5% | 45.0% | 69.2% | $0.01 | 33 sec | 4 of 5 |
Rank is among the models we ran on that review, by F1; a tie goes to the cheaper one. See the benchmark
One workspace price with seats included. Credits pool across your team. Unlimited repos on every plan.
Try AI PR review.
Get AI PR review right from day 1.
Scale AI PR review.
Custom reviewers + credits · GitHub Enterprise · SSO/SAML + audit logs · Custom security review + SLA
GPT 5.6 Luna Pro is precise and fast, and it finds few bugs: 46 of 158 known issues on our code benchmark, against 79 for Muse Spark 1.3. We have not run the other GPT 5.6 models.
We measured it as a reviewer, not as a coding agent: it reads a pull request someone else wrote and points out defects. There it was precise and found few issues, 46 of 158. How well it writes code is outside this benchmark.
Only GPT 5.6 Luna Pro, on seven of our nine review types (all but docs and cost). We have not run GPT 5.6 Sol, GPT 5.6 Terra or GPT 5.6 Luna, so none of these numbers speak for them. The Cost Review benchmark ran GPT-6 Luna Pro, a different model.
They are the same underlying model, and Luna Pro serves it in OpenAI's pro reasoning mode, which does more model work per answer and uses more tokens (OpenRouter and OpenAI, Nov 2026). Our code review run used medium effort. We have not run plain GPT 5.6 Luna, so these numbers say nothing about it.
In model fees on our benchmarks, about 12 cents a code review and 1 to 11 cents on the smaller specialist reviews. IonWarp plans include the model cost: Pro is $49/mo for the workspace.
Yes. It is one of the models IonWarp's reviewers are pinned to (DeepSeek V4.1 Flash, GPT 5.6 Luna Pro, GLM 5.3 Flash), and you can switch any reviewer's model in settings.