Measured · release 2026-10-12
GPT Luna code review, measured
GPT 5.6 Luna Pro reviewed the same pull requests as every other model on 7 of IonWarp’s 8 benchmark lanes, each finding matched against the pull request’s known issues. GPT Luna is not the best model on any lane we ran; its best F1 was 64.5%, on SOC 2 Review. On general code review, GPT 5.6 Luna Pro scored F1 41.6% at $0.116 a review; Muse Spark 1.3 led the models at 56.6%, and DeepSeek V4.1 Flash, which IonWarp’s Code Review runs, scored 49.2% at $0.102.
| Lane | Model | F1 | Rank among models | Time (p50) | $ / review |
|---|---|---|---|---|---|
| Code Review benchmark | GPT 5.6 Luna Pro | 41.6% | 5 of 5 | 1m 47s | $0.116 |
| Security Review benchmark | GPT 5.6 Luna Pro | 28.6% | 7 of 7 | 1m 16s | $0.105 |
| AI Architecture benchmark | GPT 5.6 Luna Pro | 35.3% | 4 of 8 | 51s | $0.046 |
| Log & Errors benchmark | GPT 5.6 Luna Pro | 36.0% | 5 of 5 | 44s | $0.046 |
| API Performance benchmark | GPT 5.6 Luna Pro | 50.0% | 4 of 6 | 28s | $0.030 |
| SOC 2 Review benchmark | GPT 5.6 Luna Pro | 64.5% | 3 of 5 | 37s | $0.072 |
| CI/CD Performance benchmark | GPT 5.6 Luna Pro | 54.5% | 4 of 5 | 33s | $0.014 |
Highlighted rows are what the agent runs today. Green: the best model on that lane.
FAQ
GPT Luna for review
Is GPT Luna good for code review?
GPT Luna is not the best model on any lane we ran; its best F1 was 64.5%, on SOC 2 Review. On general code review, GPT 5.6 Luna Pro scored F1 41.6% at $0.116 a review; Muse Spark 1.3 led the models at 56.6%, and DeepSeek V4.1 Flash, which IonWarp’s Code Review runs, scored 49.2% at $0.102.
What does a GPT Luna review cost?
In model charges on the benchmark, from $0.014 a review on CI/CD Performance to $0.116 on Code Review. On IonWarp, model spend comes out of the review spend your plan includes.
Which GPT Luna models has IonWarp benchmarked?
GPT 5.6 Luna Pro, on 7 of the 8 lanes: Code Review, Security Review, AI Architecture, Log & Errors, API Performance, SOC 2 Review and CI/CD Performance.
Does IonWarp run GPT Luna?
Not by default: the agents start on DeepSeek V4.1 Flash. An agent moves to another model only when IonWarpBench finds one that clears its quality bar for less, and every review names its model.
