GPT-6 Luna on BlueBench
GPT-6 Luna ranks 18th of 26 models on BlueBench, Cotool's benchmark for defensive security work, with an overall score of 64.1. It ranks highest in incident response (71.1, 13th of 26) and lowest in threat hunting (51.0, 22nd of 26). It averages $0.26 per task, the 5th lowest cost of the 26 models tested.
OpenAI · Closed weight · Updated Oct 2026
Methodology - BlueBench score
- 64.1
- #18 of 26 models
- Cost per task
- $0.26
- #5 cheapest of 26
- Reliability
- 40.6
- #10 of 26 models
- Best-ranked track
- #13
- Incident Response · 71.1
Scores by track
BlueBench weights incident response, threat hunting, and detection engineering at 30% each and malware analysis at 10%.
| Track | Score | Rank | Track leader |
|---|---|---|---|
| Incident Response | 71.1 | #13 of 26 | GPT-6 Astra · 88.0 |
| Threat Hunting | 51.0 | #22 of 26 | Claude Opus 5.5 · 76.1 |
| Detection Engineering | 69.9 | #18 of 26 | GPT-6 Astra · 87.3 |
| Malware Analysis | 65.3 | #21 of 26 | Claude Opus 5.5 · 89.4 |
Scores by benchmark
Each BlueBench benchmark is a separate intrusion, real or simulated, with its own tasks and grading.
| Benchmark | Score | Rank |
|---|---|---|
| Simulation-003Impact & Exfiltration | 68.4 | #12 of 26 |
| Simulation-002Identity & Active Directory Attacks | 55.4 | #18 of 26 |
| Simulation-001Initial Access & Command-and-Control | 39.5 | #23 of 26 |
| Intrusion-003AWS Cloud Intrusion | 78.7 | #12 of 26 |
| Intrusion-002Windows Enterprise Intrusion | 68.4 | #16 of 26 |
| Intrusion-001macOS Threat Investigation | 87.1 | #16 of 26 |
Models ranked nearby
The models closest to GPT-6 Luna on the overall leaderboard.
| Model | Rank | Score | Cost / task |
|---|---|---|---|
| #16 | 67.3 | $2.62 | |
| #17 | 65.5 | $1.03 | |
GPT-6 Luna | #18 | 64.1 | $0.26 |
| #19 | 62.7 | $3.01 | |
| #20 | 62.4 | $0.58 |
Frequently asked questions
- How good is GPT-6 Luna at security operations?
- GPT-6 Luna scores 64.1 overall on BlueBench and ranks 18th of 26 models on incident response, threat hunting, detection engineering, and malware analysis. The top model, Claude Opus 5.5, scores 83.2.
- Which security tasks is GPT-6 Luna best at?
- BlueBench track scores for GPT-6 Luna, best rank first. Incident Response: 71.1 (13th of 26); Detection Engineering: 69.9 (18th of 26); Malware Analysis: 65.3 (21st of 26); Threat Hunting: 51.0 (22nd of 26).
- How much does GPT-6 Luna cost per security task?
- GPT-6 Luna averages $0.26 per BlueBench task, including repeated attempts. That is the 5th lowest cost of 26 models. The best-scoring cheaper model is DeepSeek V4.1 Flash at $0.17 per task with an overall score of 71.5.
- How was GPT-6 Luna tested?
- Every model runs the same 74 tasks across 6 BlueBench benchmarks, built from real intrusion logs and simulated enterprise environments, using the same harness, prompts, tools, and number of trials. Reports are graded against hidden rubrics and detection rules are re-run against the evidence.
Full leaderboards
AI for the blue team.
Run Cotool's harness in your environment to get real security work done