Vals CyberBench (CyberBench)
We show this table for reference; we do not rank on it.
Can autonomous agents craft PoC inputs that trigger OSS-Fuzz vulnerabilities—and stop crashing after the fix?
Evaluation results published by Vals AI. Published leaderboard aggregates and evaluation settings only. Dataset contents are not included.
CyberBench score on CyberBench — October 12, 2026
We mirror the published cyberbench score view for CyberBench. Muse Wonder leads the public snapshot at 80.36%, followed by GPT-6 Sol (77.98%) and Gemini 4 Argon (77.86%). We do not use these results to rank models overall.
Muse Wonder
Meta
max reasoning
GPT-6 Sol
OpenAI
max reasoning
Gemini 4 Argon
high reasoning
47 modelsAgenticCurrentDisplay onlyUpdated October 12, 2026
CyberBench score table (47 models)
| Rank | Model / configuration | Score | Parameters (B) | Open / closed |
|---|---|---|---|---|
| 1 | Muse WonderMetamax reasoning | 80.36% | Not reported | Closed |
| 2 | GPT-6 SolOpenAImax reasoning | 77.98% | Not reported | Closed |
| 3 | Gemini 4 ArgonGooglehigh reasoning | 77.86% | Not reported | Closed |
| 4 | GPT-5.6 SolOpenAImax reasoning | 76.31% | Not reported | Closed |
| 5 | GPT-6 LunaOpenAImax reasoning | 76.25% | Not reported | Closed |
| 6 | Claude Haiku 5.5Anthropic | 75.48% | Not reported | Closed |
| 7 | MiMo-V2.6-FlashXiaomi | 75.36% | Not reported | Open |
| 8 | Kimi K3Moonshot AImax reasoning | 75.24% | Not reported | Pending |
| 9 | Claude Opus 5.5 RangeAnthropic | 74.58% | Not reported | Unknown |
| 10 | DeepSeek V4.1 Flash 1017DeepSeekhigh reasoning | 73.69% | Not reported | Open |
| 11 | GPT-5.6 LunaOpenAImax reasoning | 73.63% | Not reported | Closed |
| 12 | Muse Wonder FullMetamax reasoning | 73.63% | Not reported | Closed |
| 13 | GPT Daybreak BlueOpenAImax reasoning | 72.98% | Not reported | Unknown |
| 14 | DeepSeek V4 Pro 0813DeepSeekmax reasoning | 72.86% | Not reported | Open |
| 15 | MiMo-V2.6-ProXiaomi | 72.86% | Not reported | Open |
| 16 | Muse Spark 1.3 MaxMetamax reasoning | 72.74% | Not reported | Unknown |
| 17 | DeepSeek V4 Flash 0731DeepSeekhigh reasoning | 72.08% | Not reported | Open |
| 18 | GPT-5.6 TerraOpenAImax reasoning | 72.08% | Not reported | Closed |
| 19 | GLM-5.3Z.AImax reasoning | 72.08% | Not reported | Open |
| 20 | Grok 4.5xAIhigh reasoning | 71.01% | Not reported | Closed |
| 21 | Claude Fable 5.1Anthropic | 70.42% | Not reported | Closed |
| 22 | Claude Sonnet 5 RangeAnthropic | 70.30% | Not reported | Unknown |
| 23 | Muse Spark 1.2Metaxhigh reasoning | 69.52% | Not reported | Closed |
| 24 | Ember-1Fireworks | 69.52% | Not reported | Closed |
| 25 | Grok 4.7xAIxhigh reasoning | 69.46% | Not reported | Closed |
| 26 | Muse Spark 1.3Metaxhigh reasoning | 69.41% | Not reported | Closed |
| 27 | GLM-5.3-FlashZ.AImax reasoning | 67.86% | Not reported | Open |
| 28 | Qwen3.8-27BAlibabaxhigh reasoning | 67.80% | Not reported | Open |
| 29 | Grok 4.6xAIhigh reasoning | 67.68% | Not reported | Closed |
| 30 | InklingThinking Machines Lab0.99 reasoning | 66.79% | Not reported | Open |
| 31 | Claude Opus 5Anthropic | 65.36% | Not reported | Closed |
| 32 | Hy4 previewTencent | 65.36% | Not reported | Open |
| 33 | MiniMax M3MiniMax | 65.00% | Not reported | Open |
| 34 | Inkling-SmallThinking Machines Lab0.99 reasoning | 64.11% | Not reported | Open |
| 35 | Claude Sonnet 5Anthropic | 61.91% | Not reported | Closed |
| 36 | Claude Sonnet 5.5Anthropic | 59.58% | Not reported | Closed |
| 37 | Gemini 3.5 Flash-LiteGooglehigh reasoning | 59.11% | Not reported | Closed |
| 38 | Ling 3.0 Flash Af Rc3Ant | 58.33% | Not reported | Unknown |
| 39 | Claude Opus 5.5Anthropic | 55.36% | Not reported | Closed |
| 40 | Mercury 2.5Inceptionhigh reasoning | 45.48% | Not reported | Closed |
| 41 | Gemini 3.6 FlashGooglehigh reasoning | 45.36% | Not reported | Closed |
| 42 | Gemini 3.7 FlashGooglehigh reasoning | 43.75% | Not reported | Closed |
| 43 | Gemini 3.8 FlashGooglehigh reasoning | 43.75% | Not reported | Closed |
| 44 | GPT-6 AstraOpenAImax reasoning | 41.07% | Not reported | Closed |
| 45 | GPT-6.1 SolOpenAImax reasoning | 39.29% | Not reported | Closed |
| 46 | Nemotron Lightning 3p5 30b A3bFireworks AI | 29.46% | Not reported | Unknown |
| 47 | Qwen3.8 MaxAlibaba | 28.57% | Not reported | Open |
How CyberBench is shown here
BenchLM mirrors the public Vals AI CyberBench leaderboard captured from https://www.vals.ai/benchmarks/cyber and updated by Vals on October 12, 2026. The snapshot preserves overall scores, uncertainty, latency, cost-per-test metadata, and task-level scores where Vals publishes them.
CyberBench is display only on BenchLM. Vals proprietary or Vals-hosted aggregate views are useful context, but BenchLM does not use them as weighted ranking inputs or as a replacement for benchmark-native source records.
Snapshot
The published CyberBench snapshot places Muse Wonder first at 80.36%. The third row is 2.50 points behind. The broader top-10 range is 6.67 points, so many of the published results sit in a relatively narrow band.
47 models have been evaluated on CyberBench. The benchmark falls in the Agentic category. We keep external benchmark mirrors separate from the weighted global scoring system, so these results remain source-specific evidence. CyberBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About CyberBench
Year
2026
Tasks
OSS-Fuzz PoC and patch-verification tasks
Format
Accuracy score
Difficulty
Autonomous cybersecurity exploit reproduction
BenchLM mirrors the public Vals AI CyberBench leaderboard as display-only external evidence. The captured snapshot preserves overall scores, task-level scores where Vals publishes them, uncertainty, latency, and cost-per-test metadata. It is excluded from BenchLM weighted rankings.
Freshness and provenance
Version
CyberBench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
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.
Questions
What does CyberBench measure?
Can autonomous agents craft PoC inputs that trigger OSS-Fuzz vulnerabilities—and stop crashing after the fix?
Which model leads the published CyberBench snapshot?
Muse Wonder currently leads the published CyberBench snapshot with 80.36% cyberbench score. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on CyberBench?
The October 12, 2026 snapshot contains 47 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 5,600+ readers.
One email each week. Unsubscribe anytime.