Skip to main content
Radar

Every change to the models you run, with its source and its date. Releases, price changes, retirements, API changes, and incidents.Every change to the models you run, with its source.

Follow model changes
BenchLM

CyberGym

A cybersecurity task benchmark for evaluating defensive cyber workflows and vulnerability-oriented agent performance.

Data verified 33 confirmed releases in the last 30 daysSee provider release alerts

Benchmark score on CyberGym — September 15, 2026

We mirror the published score view for CyberGym. DeepSeek V4.1 Flash leads the public snapshot at 88.1%, followed by Fugu Cyber (86.9%) and Atria Dawn Preview (86.5%). We do not use these results to rank models overall.

1Open weights

DeepSeek V4.1 Flash

DeepSeek

88.1%
Context 1M
2Proprietary

Fugu Cyber

Sakana AI

86.9%
Context 1M
3Open weights

Atria Dawn Preview

Shanghai Artificial Intelligence Laboratory

86.5%
Context 256K

24 modelsAgenticCurrentDisplay onlyUpdated September 15, 2026

Benchmark score table (24 models)

Score
1
DeepSeek V4.1 FlashDeepSeek · Open weight
88.1%
2
Fugu CyberSakana AI · Closed
86.9%
3
Atria Dawn PreviewShanghai Artificial Intelligence Laboratory · Open weight
86.5%
4
Gemini 3.8 Flash CyberGoogle · Closed
86.2%
5
GPT-5.6 SolOpenAI · Closed
84.5%
6
GLM-5.3Z.AI · Open weight
84.5%
7
Claude Mythos 5Anthropic · Closed
83.8%
8
DeepSeek V4 Pro 0813DeepSeek · Closed
83.3%
9
Gemini 3.5 Flash CyberGoogle · Closed
83.2%
10
Claude Mythos PreviewAnthropic · Closed
83.1%
11
GPT-5.6 TerraOpenAI · Closed
81.8%
12
GPT-5.5OpenAI · Closed
81.8%
13
GPT-5.4OpenAI · Closed
79.0%
14
Hy4 previewTencent · Open weight
78.4%
15
GPT-5.6 LunaOpenAI · Closed
77.9%
16
DeepSeek V4 Flash 0731DeepSeek · Closed
76.7%
17
Claude Opus 4.7 (Adaptive)Anthropic · Closed
73.1%
18
GLM-5.1Z.AI · Open weight
68.7%
19
Claude Opus 4.6Anthropic · Closed
66.6%
20
Claude Sonnet 4.6Anthropic · Closed
65.2%
21
Muse Spark 1.1Meta · Closed
59.0%
22
Claude Opus 4.5Anthropic · Closed
50.6%
23
Muse SparkMeta · Closed
43.5%
24
GLM-5Z.AI · Open weight
43.2%

The published CyberGym snapshot places DeepSeek V4.1 Flash first at 88.1%. The third row is 1.6 points behind. The broader top-10 range is 5.0 points, so many of the published results sit in a relatively narrow band.

24 models have been evaluated on CyberGym. The benchmark falls in the Agentic category. This category carries a 22% weight in BenchLM.ai's overall scoring system. CyberGym is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About CyberGym

Year

2026

Tasks

1,507 vulnerability analysis instances

Format

Vulnerability reproduction and PoC generation

Difficulty

Real-world cybersecurity

CyberGym includes 1,507 benchmark instances from historical vulnerabilities across 188 large software projects. BenchLM stores CyberGym as a display-only agentic security benchmark when exact provider comparison values are published.

CyberGym: Evaluating AI Agents' Real-World Cybersecurity Capabilities at Scale

BenchLM freshness & provenance

Version

CyberGym 2026

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark set

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

FAQ

What does CyberGym measure?

A cybersecurity task benchmark for evaluating defensive cyber workflows and vulnerability-oriented agent performance.

Which model scores highest on CyberGym?

DeepSeek V4.1 Flash by DeepSeek currently leads with a score of 88.1% on CyberGym.

How many models are evaluated on CyberGym?

24 AI models have been evaluated on CyberGym on BenchLM.

Compare Top Models on CyberGym

DeepSeek V4.1 Flash vs Fugu CyberFugu Cyber vs Atria Dawn PreviewAtria Dawn Preview vs Gemini 3.8 Flash CyberGemini 3.8 Flash Cyber vs GPT-5.6 Sol
Last updated: September 15, 2026 · BenchLM version CyberGym 2026

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 2,000+ readers.

One email each week. Unsubscribe anytime.