Try for free

Measured · release 2026-10-12

The best LLM for code review, measured

Muse Spark 1.3 found 79 of 158 known issues on 50 pull requests, with 42 findings that matched none: F1 56.6%, for $0.424 a review. DeepSeek V4.1 Flash, which IonWarp’s Code Review runs by default, scored 49.2% for $0.102. Claude Opus 4.5 graded every finding.

Code review, 50 PRs: models and appsRanked
Code review, 50 PRs: models and apps
Model / AppReviewer ScoreF1Latency$/run
IonWarp · Muse Spark 1.371.256.6%3m 18s$0.424
Qodo v264.758.5%5m 40s~$1.67
IonWarp · DeepSeek V4 Pro48.751.4%5m 59s$0.165
IonWarp · DeepSeek V4.1 FlashRuns on your PRs41.049.2%6m 56s$0.102
IonWarp · GLM 5.3 Flash40.842.8%2m 15s$0.018
Greptile v4.139.951.1%4m 58s~$1.75
Cursor Bugbot36.251.4%7m 52s~$1.25
IonWarp · GPT 5.6 Luna Pro31.841.6%1m 47s$0.116
CodeRabbit3.142.1%7m 34s~$1.50

Reviewer Score = (3 × F1 + time + $/run) ÷ 5, each part 0 to 100 within this table. IonWarp rows cost the model’s inference; ~$ is each app’s list price for 5 seats at 100 reviews a month. Pinned Martian Core reproduction. Release 2026-10-12, CC BY 4.0.

By kind of review

The best model for each review

The top F1 among models on each lane. What wins general code review is not what wins security or CI/CD.

F1 by model and lanegreen: best model on the lane · outlined: what the agent runs
F1 by model and lane
ModelCode Review benchmarkSecurity Review benchmarkAI Architecture benchmarkLog & Errors benchmarkDocs Review benchmarkAPI Performance benchmarkSOC 2 Review benchmarkCI/CD Performance benchmark
DeepSeek V4 Pro51.4%34.8%44.4%–65.2%52.2%56.3%72.7%
DeepSeek V4.1 Flash49.2%62.5%47.1%59.0%–53.3%66.7%59.5%
GLM 5.3––47.1%–––––
GLM 5.3 Flash42.8%66.7%15.4%36.1%–48.3%66.7%55.0%
GPT 5.6 Luna Pro41.6%28.6%35.3%36.0%–50.0%64.5%54.5%
Gemini 3.8 Flash–51.6%33.3%67.7%68.6%41.7%––
Grok 4.6–53.3%34.8%–––––
Muse Spark 1.356.6%44.4%28.6%43.6%59.5%56.0%55.2%44.4%

Models only, each run as IonWarp’s reviewer for that lane on every pull request of the lane. Only Code Review is ranked; the other lanes are provisional. On a phone this shows the first 3 of 8 lanes; a wider screen shows them all.

Model by model

Where each model leads

Muse Spark 1.3

Muse Spark

Best model on Code Review, API Performance. Best F1: 59.5% on Docs Review; lowest: 28.6% on AI Architecture.

Muse Spark, lane by lane
DeepSeek V4 Pro · DeepSeek V4.1 Flash

DeepSeek

Best model on AI Architecture, CI/CD Performance. Best F1: 72.7% on CI/CD Performance; lowest: 34.8% on Security Review.

DeepSeek, lane by lane
GLM 5.3 Flash · GLM 5.3

GLM

Best model on Security Review, SOC 2 Review. Best F1: 66.7% on Security Review; lowest: 15.4% on AI Architecture.

GLM, lane by lane
GPT 5.6 Luna Pro

GPT Luna

Not the best model on any lane. Best F1: 64.5% on SOC 2 Review; lowest: 28.6% on Security Review.

GPT Luna, lane by lane
Grok 4.6

Grok

Not the best model on any lane. Best F1: 53.3% on Security Review; lowest: 34.8% on AI Architecture.

Gemini 3.8 Flash

Gemini

Best model on Log & Errors, Docs Review. Best F1: 68.6% on Docs Review; lowest: 33.3% on AI Architecture.

Gemini, lane by lane

FAQ

Picking a model

What is the best LLM for code review?

On 50 pull requests with 158 known issues, Muse Spark 1.3 had the best F1 of the 5 models we ran: 56.6%, at $0.424 a review. The Qodo v2 app scored 58.5%. DeepSeek V4.1 Flash, which IonWarp’s Code Review runs, scored 49.2% at $0.102.

Is the best model the same for every kind of review?

No. Code Review: Muse Spark 1.3; Security Review: GLM 5.3 Flash; AI Architecture: DeepSeek V4.1 Flash; Log & Errors: Gemini 3.8 Flash; Docs Review: Gemini 3.8 Flash; API Performance: Muse Spark 1.3; SOC 2 Review: GLM 5.3 Flash; CI/CD Performance: DeepSeek V4 Pro. Only the code review lane is ranked; the others are provisional.

Why doesn’t IonWarp run the top model everywhere?

Cost and speed count too. Every agent starts on DeepSeek V4.1 Flash, and IonWarpBench moves an agent to another model only when it clears the quality bar for less. The model is named on every review.

Is the best LLM for coding also the best for code review?

Not necessarily, and a coding ranking cannot tell you. Coding benchmarks grade code a model writes; this one grades whether a model finds known issues in a pull request someone else wrote.

Every agent names its model.

Try for free

14 days free, no card. Then Pro is $49 a month per workspace with 5 seats, or Max $149 with 10.