Skip to main content

Kimi K3 security benchmark rundown

Kimi K3 (Moonshot AI) ranks 16th of 26 on BlueBench with 67.3 overall at $2.62 per task. Track scores, value against peers, strengths, weaknesses, and when to use it.

Moonshot AI · Open weight · Tested Sep 11, 2026 · Data updated Oct 2026
Methodology

Verdict

Kimi K3 ranks 16th of 26 on BlueBench with an overall score of 67.3, in the lower half of the leaderboard and 15.9 points behind Claude Opus 5.5. At $2.62 per task it is the 17th cheapest of 26, and Claude Sonnet 5.5 scores higher (77.5) for less ($1.01). Against the 10 models priced within 2x of its cost per task, it scores 3.7 points below the peer median overall; it trails them most in detection engineering.

BlueBench score
67.3
#16 of 26 models
Cost per task
$2.62
#17 cheapest of 26
Reliability
36.3
#15 of 26 models
Overall vs cost peers
−3.7
Off the cost–score frontier
All Kimi K3 results and FAQ →

Capability by track

Incident response, threat hunting, and detection engineering each carry 30% of the BlueBench score; malware analysis carries 10%.

Kimi K3 scores 67.3 overall, 16th of 26 on the capability leaderboard. Its best rank is in threat hunting (64.3, 8th of 26) and its worst is in detection engineering (62.0, 24th of 26). Among the 8 open-weight models tested it ranks 4th.

TrackWeightScoreRankPeer medianvs peers
Incident Response30%72.3#12 of 2672.8−0.5
Threat Hunting30%64.3#8 of 2659.5+4.8
Detection Engineering30%62.0#24 of 2679.8−17.7
Malware Analysis10%77.2#13 of 2681.1−3.9

Cost, reliability, and value

Cost is USD per evaluated task including repeated attempts. Reliability is 100 minus the mean score range across the six benchmarks.

Kimi K3 averages $2.62 per task, including repeated attempts: the 17th cheapest of 26 models with a cost figure. It is off the cost–score frontier: Claude Sonnet 5.5 scores 77.5 at $1.01, and 8 other models are also both cheaper and at least as good. Its peer group is the 10 models priced within 2x of its cost per task. Their median overall score is 71.0, so Kimi K3 is 3.7 points behind its price range. Reliability, 100 minus its mean score range across the benchmarks, is 36.3 (15th of 26).

Strengths and weaknesses

Each track compared with the median of the 10 models priced within 2x of its cost per task. A gap of 5 points or more counts.

Against the 10 models priced within 2x of its cost per task, Kimi K3 lags in detection engineering (−17.7). Differences are versus the peer median score in each track; ±5 points is the threshold.

Beats its cost peers

No track by 5+ points

Lags its cost peers

  • Detection Engineering62.0 (−17.7)

Nearest neighbors on the leaderboard

The models ranked just above and below it overall, with the largest track gaps.

  • GPT-5.6 Terra sits 2 places above, 2.6 points higher overall (69.9), at 38% of the cost per task ($1.01); it is 15.4 points ahead in detection engineering.
  • Claude Haiku 5.5 sits one place above, 1.2 points higher overall (68.5), at 23% of the cost per task ($0.59); it is 18.5 points ahead in detection engineering and 9.4 points behind in malware analysis.
  • GLM-5.2 sits one place below, 1.8 points lower overall (65.5), at 39% of the cost per task ($1.03); it is 13.9 points ahead in detection engineering and 11.9 points behind in threat hunting.
  • GPT-6 Luna sits 2 places below, 3.2 points lower overall (64.1), at 10% of the cost per task ($0.26); it is 7.9 points ahead in detection engineering and 13.3 points behind in threat hunting.
RankModelScore$/task
14GPT-5.6 Terra69.9$1.01
15Claude Haiku 5.568.5$0.59
16Kimi K367.3$2.62
17GLM-5.265.5$1.03
18GPT-6 Luna64.1$0.26

Best for

Alternatives are limited to models costing at most 1.5x as much, except the overall leader and the cheaper-model pick.

  • Use itBlueBench does not single out a task for it: it is in the top 3 of no track and beats its cost peers by 5+ points in none.
  • InsteadUse Claude Sonnet 5.5 instead in most cases: it scores 77.5 overall (+10.2) at $1.01 per task (39% of the cost per task).
  • InsteadUse Claude Opus 4.8 instead if detection engineering is the priority: it scores 85.3 there (vs 62.0) at $1.56 per task.
  • InsteadUse Claude Opus 5.5 instead if the highest overall score matters more than cost: it scores 83.2 (15.9 points higher) at $3.54 per task.

Notable benchmark results

Each BlueBench benchmark is a separate intrusion, real or simulated, with its own tasks and grading.

  • Its best standing is on Intrusion-002 (Windows Enterprise Intrusion): 76.0 (7th of 26).
  • Its scores range from 57.5 on Simulation-001 to 79.9 on Intrusion-003, a spread of 22.4 points.
BenchmarkScoreRank
Simulation-003Impact & Exfiltration62.8#18 of 26
Simulation-002Identity & Active Directory Attacks60.2#16 of 26
Simulation-001Initial Access & Command-and-Control57.5#8 of 26
Intrusion-003AWS Cloud Intrusion79.9#10 of 26
Intrusion-002Windows Enterprise Intrusion76.0#7 of 26
Intrusion-001macOS Threat Investigation75.1#22 of 26

About this rundown

Provider
Moonshot AI · open weight
Tested
Sep 11, 2026
Suite
74 tasks across 6 benchmarks, evaluated in a common harness
Data updated
Oct 27, 2026
How it is written
Generated from the BlueBench result files with fixed rules: cost peers are models priced within 2x per task (or the nearest by cost when fewer than three are), a track is a strength or weakness at 5+ points from the peer median, and ranks share a place on ties. The text is regenerated whenever the results change.

AI for the blue team.

Run Cotool's harness in your environment to get real security work done

Book a demo