Benchmark Almanac

CyberBench (Vals)

GPT-5.6 Sol 88.1% leads since July 2026
Not in model cards

Agents work on OSS-Fuzz memory-safety bugs from ARVO images in two tracks. They write a proof-of-concept input that crashes the target, or patch the source so the crash stops and tests still pass.

The PoC track follows CyberGym's method. Runs used mini-SWE-agent. Not the 2024 CyberBench by Liu et al. Archived after v1.1 replaced it.

Made by
Vals AI
Released
June 2026
Kind
Finding and exploiting real bugs
Level
Expert
Scored
Checked automatically
Answers
Agent task
Tasks
Kept private
Strongest model tested
Claude Opus 5 #10 on the capability index
Last model added
September 2026
Scores refreshed
November 12th, 2026
26
Models scored on its own leaderboard

GPT-5.6 Sol leads with 88.1% among models released by September 2026

Best score per model from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.

The newest model with a score here came out in September 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

26 models. Scroll sideways for the rest.

  1. 88.1 GPT-5.6 Sol
  2. 83.9 GPT-5.6 Luna
  3. 80.5 GPT-5.5
  4. 79.0 Kimi K3
  5. 78.8 DeepSeek V4.1 Flash
  6. 77.5 DeepSeek V4 Flash 0731
  7. 77.4 GLM-5.2
  8. 76.3 MiniMax-M3
  9. 74.3 Kimi K2.6
  10. 72.9 GPT-5.4
  11. 72.3 Gemini 3.5 Flash-Lite
  12. 71.5 GLM-5.3-Flash
  13. 69.3 Gemini 3.5 Flash
  14. 68.8 Inkling-Small
  15. 68.6 DeepSeek-V4-Pro
  16. 66.7 DeepSeek-V4-Flash
  17. 64.1 Inkling
  18. 57.5 Tencent Hy4 preview
  19. 52.1 Claude Opus 4.7
  20. 50.9 Claude Opus 4.8
  21. 48.5 Gemini 3.6 Flash
  22. 47.5 Grok 4.3
  23. 43.8 Gemini 3.8 Flash
  24. 40.7 Claude Opus 5
  25. 37.5 Qwen3.7-Plus
  26. 36.4 Gemini 3.1 Pro
  1. 1 GPT-5.6 Sol 88.1%
  2. 2 GPT-5.6 Luna 83.9%
  3. 3 GPT-5.5 80.5%
  4. 4 Kimi K3 79%
  5. 5 DeepSeek V4.1 Flash 78.8%
  6. 6 DeepSeek V4 Flash 0731 77.5%
  7. 7 GLM-5.2 77.4%
  8. 8 MiniMax-M3 76.3%
  9. 9 Kimi K2.6 74.3%
  10. 10 GPT-5.4 xhigh 72.9%
  11. 11 Gemini 3.5 Flash-Lite 72.3%
  12. 12 GLM-5.3-Flash 71.5%
  13. 13 Gemini 3.5 Flash 69.3%
  14. 14 Inkling-Small 68.8%
  15. 15 DeepSeek-V4-Pro 68.6%
  16. 16 DeepSeek-V4-Flash 66.7%
  17. 17 Inkling 64.1%
  18. 18 Tencent Hy4 preview 57.5%
  19. 19 Claude Opus 4.7 52.1%
  20. 20 Claude Opus 4.8 50.9%
  21. 21 Gemini 3.6 Flash 48.5%
  22. 22 Grok 4.3 47.5%
  23. 23 Gemini 3.8 Flash 43.8%
  24. 24 Claude Opus 5 40.7%
  25. 25 Qwen3.7-Plus 37.5%
  26. 26 Gemini 3.1 Pro 36.4%

Source: Official CyberBench (Vals) leaderboard

#ModelScoreSettingDateSource
1 GPT-5.6 Sol 88.1% Jul 2026 Official leaderboard
2 GPT-5.6 Luna 83.9% Jul 2026 Official leaderboard
3 GPT-5.5 80.5% Apr 2026 Official leaderboard
4 Kimi K3 79% Jul 2026 Official leaderboard
5 DeepSeek V4.1 Flash 78.8% Sep 2026 Official leaderboard
6 DeepSeek V4 Flash 0731 77.5% Jul 2026 Official leaderboard
7 GLM-5.2 77.4% Jun 2026 Official leaderboard
8 MiniMax-M3 76.3% Jun 2026 Official leaderboard
9 Kimi K2.6 74.3% Apr 2026 Official leaderboard
10 GPT-5.4 72.9% xhigh Mar 2026 Official leaderboard
11 Gemini 3.5 Flash-Lite 72.3% Jul 2026 Official leaderboard
12 GLM-5.3-Flash 71.5% Aug 2026 Official leaderboard
13 Gemini 3.5 Flash 69.3% May 2026 Official leaderboard
14 Inkling-Small 68.8% Jul 2026 Official leaderboard
15 DeepSeek-V4-Pro 68.6% Apr 2026 Official leaderboard
16 DeepSeek-V4-Flash 66.7% Apr 2026 Official leaderboard
17 Inkling 64.1% Jul 2026 Official leaderboard
18 Tencent Hy4 preview 57.5% Aug 2026 Official leaderboard
19 Claude Opus 4.7 52.1% Apr 2026 Official leaderboard
20 Claude Opus 4.8 50.9% May 2026 Official leaderboard
21 Gemini 3.6 Flash 48.5% Jul 2026 Official leaderboard
22 Grok 4.3 47.5% Apr 2026 Official leaderboard
23 Gemini 3.8 Flash 43.8% Sep 2026 Official leaderboard
24 Claude Opus 5 40.7% Jul 2026 Official leaderboard
25 Qwen3.7-Plus 37.5% Jun 2026 Official leaderboard
26 Gemini 3.1 Pro 36.4% Feb 2026 Official leaderboard

Accuracy against latency

The Vals scatter chart with its x axis switched from Cost to Latency: mean time per test against accuracy, on a log scale. The line joins the runs that no other run beats on both measures. We redrew this chart from the numbers on the official leaderboard.

40% 60% 80% 100% 1001,00010K Latency (mean per test) (s), log scale, lower is better Accuracy Gemini 3.1 Pro: Accuracy 36.4%, latency (mean per test) 1,032 s GPT-5.4 (xhigh): Accuracy 72.9%, latency (mean per test) 1,109 s Claude Opus 4.7: Accuracy 52.1%, latency (mean per test) 752 s Kimi K2.6: Accuracy 74.3%, latency (mean per test) 1,187 s GPT-5.5: Accuracy 80.5%, latency (mean per test) 680 s DeepSeek-V4-Flash: Accuracy 66.7%, latency (mean per test) 858 s DeepSeek-V4-Pro: Accuracy 68.6%, latency (mean per test) 1,120 s Grok 4.3: Accuracy 47.5%, latency (mean per test) 1,884 s Gemini 3.5 Flash: Accuracy 69.3%, latency (mean per test) 962 s Claude Opus 4.8: Accuracy 50.9%, latency (mean per test) 897 s MiniMax-M3: Accuracy 76.3%, latency (mean per test) 1,434 s GLM-5.2: Accuracy 77.4%, latency (mean per test) 1,349 s GPT-5.6 Luna: Accuracy 83.9%, latency (mean per test) 575 s Inkling: Accuracy 64.1%, latency (mean per test) 1,999 s Inkling-Small: Accuracy 68.8%, latency (mean per test) 2,843 s Kimi K3: Accuracy 79%, latency (mean per test) 1,153 s Gemini 3.5 Flash-Lite: Accuracy 72.3%, latency (mean per test) 1,447 s Gemini 3.6 Flash: Accuracy 48.5%, latency (mean per test) 1,442 s Claude Opus 5: Accuracy 40.7%, latency (mean per test) 2,022 s DeepSeek V4 Flash 0731: Accuracy 77.5%, latency (mean per test) 1,053 s GLM-5.3-Flash: Accuracy 71.5%, latency (mean per test) 2,924 s Tencent Hy4 preview: Accuracy 57.5%, latency (mean per test) 2,291 s Gemini 3.8 Flash: Accuracy 43.8%, latency (mean per test) 621 s DeepSeek V4.1 Flash: Accuracy 78.8%, latency (mean per test) 758 s Qwen3.7-Plus: Accuracy 37.5%, latency (mean per test) 454 s GPT-5.6 Sol: Accuracy 88.1%, latency (mean per test) 537 s GPT-5.6 SolQwen3.7-Plus 40% 60% 80% 100% 1001,00010K Latency (mean per test) (s), log scale, lower is better Accuracy Gemini 3.1 Pro: Accuracy 36.4%, latency (mean per test) 1,032 s GPT-5.4 (xhigh): Accuracy 72.9%, latency (mean per test) 1,109 s Claude Opus 4.7: Accuracy 52.1%, latency (mean per test) 752 s Kimi K2.6: Accuracy 74.3%, latency (mean per test) 1,187 s GPT-5.5: Accuracy 80.5%, latency (mean per test) 680 s DeepSeek-V4-Flash: Accuracy 66.7%, latency (mean per test) 858 s DeepSeek-V4-Pro: Accuracy 68.6%, latency (mean per test) 1,120 s Grok 4.3: Accuracy 47.5%, latency (mean per test) 1,884 s Gemini 3.5 Flash: Accuracy 69.3%, latency (mean per test) 962 s Claude Opus 4.8: Accuracy 50.9%, latency (mean per test) 897 s MiniMax-M3: Accuracy 76.3%, latency (mean per test) 1,434 s GLM-5.2: Accuracy 77.4%, latency (mean per test) 1,349 s GPT-5.6 Luna: Accuracy 83.9%, latency (mean per test) 575 s Inkling: Accuracy 64.1%, latency (mean per test) 1,999 s Inkling-Small: Accuracy 68.8%, latency (mean per test) 2,843 s Kimi K3: Accuracy 79%, latency (mean per test) 1,153 s Gemini 3.5 Flash-Lite: Accuracy 72.3%, latency (mean per test) 1,447 s Gemini 3.6 Flash: Accuracy 48.5%, latency (mean per test) 1,442 s Claude Opus 5: Accuracy 40.7%, latency (mean per test) 2,022 s DeepSeek V4 Flash 0731: Accuracy 77.5%, latency (mean per test) 1,053 s GLM-5.3-Flash: Accuracy 71.5%, latency (mean per test) 2,924 s Tencent Hy4 preview: Accuracy 57.5%, latency (mean per test) 2,291 s Gemini 3.8 Flash: Accuracy 43.8%, latency (mean per test) 621 s DeepSeek V4.1 Flash: Accuracy 78.8%, latency (mean per test) 758 s Qwen3.7-Plus: Accuracy 37.5%, latency (mean per test) 454 s GPT-5.6 Sol: Accuracy 88.1%, latency (mean per test) 537 s GPT-5.6 SolQwen3.7-Plus
ModelSettingAccuracyLatency (mean per test)
GPT-5.6 Sol 88.1% 536.99 s
GPT-5.6 Luna 83.9% 574.9 s
GPT-5.5 80.5% 680.05 s
Kimi K3 79% 1,153 s
DeepSeek V4.1 Flash 78.8% 757.63 s
DeepSeek V4 Flash 0731 77.5% 1,053 s
GLM-5.2 77.4% 1,349 s
MiniMax-M3 76.3% 1,434 s
Kimi K2.6 74.4% 1,187 s
GPT-5.4 xhigh 72.9% 1,109 s
Gemini 3.5 Flash-Lite 72.3% 1,447 s
GLM-5.3-Flash 71.5% 2,924 s
Gemini 3.5 Flash 69.3% 961.94 s
Inkling-Small 68.8% 2,843 s
DeepSeek-V4-Pro 68.6% 1,120 s
DeepSeek-V4-Flash 66.7% 857.6 s
Inkling 64.1% 1,999 s
Tencent Hy4 preview 57.5% 2,291 s
Claude Opus 4.7 52.1% 752.38 s
Claude Opus 4.8 50.9% 896.54 s
Gemini 3.6 Flash 48.5% 1,442 s
Grok 4.3 47.5% 1,884 s
Gemini 3.8 Flash 43.8% 620.68 s
Claude Opus 5 40.7% 2,022 s
Qwen3.7-Plus 37.5% 454.08 s
Gemini 3.1 Pro 36.4% 1,032 s

The best score, 88.1%, costs $1.15 a task

Runs on the cost frontier score higher than any run that's cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.

OpenAI DeepSeek Cost frontier One model at each effort level
0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 Cost per task, log scale GPT-5.6 Sol DeepSeek V4.1 Flash 0% 20% 40% 60% 80% 100% $0.01 $0.10 $1.00 $10 Cost per task, log scale GPT-5.6 Sol DeepSeek V4.1 Flash

Source: Official CyberBench (Vals) leaderboard

ModelEffortHarnessScoreCost per taskOutput tokens
GPT-5.6 Sol 88.1% $1.15
GPT-5.6 Luna 83.9%
GPT-5.5 80.5% $2.81
Kimi K3 79% $1.48
DeepSeek V4.1 Flash 78.8% $0.07
DeepSeek V4 Flash 0731 77.5% $0.08
GLM-5.2 77.4% $1.21
MiniMax-M3 76.3% $1.74
Kimi K2.6 74.3% $1.05
GPT-5.4 xhigh 72.9% $2.95
Gemini 3.5 Flash-Lite 72.3% $0.55
GLM-5.3-Flash 71.5% $0.11
Gemini 3.5 Flash 69.3% $2.00
Inkling-Small 68.8% $0.19
DeepSeek-V4-Pro 68.6% $0.68
DeepSeek-V4-Flash 66.7% $0.20
Inkling 64.1% $4.85
Tencent Hy4 preview 57.5% $0.49
Claude Opus 4.7 52.1% $4.09
Claude Opus 4.8 50.9% $1.71
Gemini 3.6 Flash 48.5% $0.87
Grok 4.3 47.5% $2.56
Gemini 3.8 Flash 43.8% $1.31
Claude Opus 5 40.7% $2.87
Qwen3.7-Plus 37.5% $0.10
Gemini 3.1 Pro 36.4% $2.10

2 models have held number one on the official leaderboard since its release in June 2026

Lab claims count from the day the card came out, independent runs from the model’s release.

OpenAI
Own leaderboard Passed 85% in Jul 2026, by GPT-5.6 Sol
GPT-5.6 Sol 88.1%

Source: Official CyberBench (Vals) leaderboard

GPT-5.6 Sol on the official leaderboard passed 85% in July 2026

All the scores we have, one dot each, so a model tested at several settings appears more than once.

On the benchmark’s own leaderboard Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) GPT-5.6 Sol GPT-5.5 Kimi K2.6 Gemini 3.1 Pro 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) GPT-5.6 Sol GPT-5.5

Source: Official CyberBench (Vals) leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

Versions

We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.

OpenAI Meta
CyberBench v1.1 (Vals) 4 leaders
Muse Wonder 80.4%
CyberBench (Vals) Passed 85% in Jul 2026, by GPT-5.6 Sol
GPT-5.6 Sol 88.1%
CyberBench v1.1 (Vals) 2 leaders
Muse Wonder 80.4%

Source: Official leaderboards

Replaced by CyberBench v1.1 (Vals)

More in Cybersecurity

All 115 in Cybersecurity
  • ExploitGym In use 869 real-vulnerability exploit-development tasks (userspace, V8, Linux kernel). Success means capturing a flag through the intended vulnerability. GPT-6 Astra 42.4%
  • ExploitBench In use 41 recent V8 JavaScript-engine N-day vulnerabilities. Progress measured by 16 grader-verified capability flags across five tiers. GPT-6 Astra 100%
  • SRE-Bench In use Software reverse engineering without source: 262 binary instances from 19 private programs, each with six graded objectives. GPT-6 Astra 56.9%
  • CyScenarioBench In use Scenario-based long-horizon security challenges in sandboxes, organized in task, scenario and campaign tiers. Evaluation set kept private. Gemini 4 Argon 100%
  • ExploitBench - Internal Port In use Agents develop working exploits for recently disclosed vulnerabilities to limit contamination. Scored as arbitrary code execution success rate. GPT-6 Cyber 42%
  • MiMo Cyber Bench In use Agents do in-house cybersecurity tasks that add to public vulnerability reproduction and exploitation benchmarks. The report gives no task count or grading details. MiMo-V2.6 Pro 81.7%