Benchmark Almanac

CyScenarioBench

Gemini 4 Argon 100% leads since Oct 2026 Only in system cards

Scenario-based long-horizon security challenges in sandboxes, organized in task, scenario and campaign tiers. Evaluation set kept private.

Scenario based evaluation of long-horizon offensive operations, scored stage by stage. Difficulty is rated with Irregular's SOLVE+ scoring system, which grades reconnaissance, OpSec, social engineering and planning against human skill levels. Irregular reports counts of 11 and 13 scenarios in different 2026 assessments, so the set grows. Not public.

Made by Irregular Released Dec 2025 Capture-the-flag challenges Tasks kept private Last in a system card Oct 2026 Last model added Oct 2026
Level
Expert Specialists with years of training, often over hours of work
Scored
Checked automatically A program checks the result against an answer key, unit tests or the expected final state.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
Paper Leaderboard irregular.com
14
Models scored in lab cards
15
System cards report it, from 4 labs

Gemini 4 Argon leads with 100%

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

14 models. Scroll sideways for the rest.

  1. 100.0 Gemini 4 Argon
  2. 67.6 Claude Opus 5.5
  3. 59.0 GPT-6 Astra
  4. 46.1 Claude Sonnet 5.5
  5. 33.7 Claude Opus 5
  6. 28.0 GPT-5.6 Sol
  7. 26.0 GPT-5.5
  8. 11.0 GPT-5.4
  9. 3.30 Claude Haiku 5.5
  10. 2.00 Muse Spark 1.3
  11. 0.50 Muse Spark 1.1
  12. 0.0 GPT-5.2 Codex
  13. 0.0 GPT-5.3 Codex
  14. 0.0 Muse Spark
  1. 1 Gemini 4 Argon 100%
  2. 2 Claude Opus 5.5 67.6%
  3. 3 GPT-6 Astra 59%
  4. 4 Claude Sonnet 5.5 46.1%
  5. 5 Claude Opus 5 33.7%
  6. 6 GPT-5.6 Sol 28%
  7. 7 GPT-5.5 26%
  8. 8 GPT-5.4 11%
  9. 9 Claude Haiku 5.5 3.30%
  10. 10 Muse Spark 1.3 2.00%
  11. 11 Muse Spark 1.1 0.50%
  12. 12 GPT-5.2 Codex 0%
  13. 13 GPT-5.3 Codex 0%
  14. 14 Muse Spark 0%
#ModelScoreSettingDateSource
1 Gemini 4 Argon 100% Oct 2026 Model card
2 Claude Opus 5.5 67.6% Sep 2026 Model card
3 GPT-6 Astra 59% Sep 2026 Model card
4 Claude Sonnet 5.5 46.1% Sep 2026 Model card
5 Claude Opus 5 33.7% Jul 2026 Model card
6 GPT-5.6 Sol 28% Jun 2026 Model card
7 GPT-5.5 26% Apr 2026 Model card
8 GPT-5.4 11% Mar 2026 Model card
9 Claude Haiku 5.5 3.30% Oct 2026 Model card
10 Muse Spark 1.3 2.00% Oct 2026 Model card
11 Muse Spark 1.1 0.50% Jul 2026 Model card
12 GPT-5.2 Codex 0% Dec 2025 Model card
13 GPT-5.3 Codex 0% Feb 2026 Model card
14 Muse Spark 0% Jun 2026 Model card

8 models have held number one since its release in Dec 2025

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

OpenAI Anthropic Google DeepMind
Lab claims Passed 85% in Oct 2026, by Gemini 4 Argon
Gemini 4 Argon 100%

The best score passed 85% in Oct 2026

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

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Gemini 4 Argon Claude Opus 5.5 GPT-6 Astra Claude Opus 5 GPT-5.6 Sol GPT-5.4 0% 20% 40% 60% 80% 100% 2026 2027 Near the ceiling (85%) Gemini 4 Argon Claude Opus 5.5

Labs report it only in system cards

We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark. Some system cards do, below.

15 system cards report it, from 4 labs

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
23 Oct 2026 Muse Spark 1.3 system card 2 CyScenarioBench pass@1, all 10 scenarios, Muse Spark 1.3 (max) (Table 1, Table 4). pass@20 = 30.0, 3 of 10 solved Cyber risk
13 Oct 2026 Gemini 4 Argon system card 100 Irregular CyScenarioBench, challenges solved end-to-end 12/12; segments passed 77/82 (93.9%) (Figure 3.2, 3.2.2.3) Cyber risk
7 Oct 2026 Claude Haiku 5.5 system card 3.3 CyScenarioBench, overall solve rate on the 10-challenge subset, pre-release snapshot, cyber mitigations off (3.3.2) Cyber risk
28 Sep 2026 Claude Sonnet 5.5 system card 46.1 CyScenarioBench, average solve rate over a 10-challenge subset, cyber mitigations off (3.2.2, Figure 3.2.2.A) Cyber risk
22 Sep 2026 Claude Opus 5.5 system card 67.6 CyScenarioBench, overall solve rate averaged over a 10-challenge subset, cyber mitigations off (3.3.2, Figure 3.3.2.A) Cyber risk
3 Sep 2026 GPT-6 Astra system card 59 Irregular CyScenarioBench, average success rate (9 of 10 solved at least once) (10.1.2.6) Cyber risk
24 Jul 2026 Claude Opus 5 system card 33.7 CyScenarioBench, overall solve rate averaged over a 9-challenge subset, mitigations off (3.3.4) Cyber risk
9 Jul 2026 GPT-5.6 Sol system card 28 Irregular CyScenarioBench, average success, 7 of 11 challenges solved (9.1.2.5) Cyber risk
9 Jul 2026 Muse Spark 1.1 system card 0.5 CyScenarioBench pass@1 (Table 1, Figure 11). pass@20 = 10.0, 1 of 10 scenarios Cyber risk
25 Jun 2026 GPT-5.6 Sol system card 28 Irregular CyScenarioBench, average success, 7 of 11 challenges solved (9.1.2.5) Cyber risk
12 Jun 2026 Muse Spark system card 0 CyScenarioBench, 10 scenarios (Table 1) Cyber risk
23 Apr 2026 GPT-5.5 system card 26 Irregular CyScenarioBench, average success rate, 7 of 11 challenges solved (9.1.2.5) Cyber risk
5 Mar 2026 GPT-5.4 system card 11 Irregular CyScenarioBench, average success rate, 5 of 11 challenges solved (5.1.2.4) Cyber risk
5 Feb 2026 GPT-5.3 Codex system card 0 CyScenarioBench by Irregular, text says no challenges solved (5.1.2.4) Cyber risk
18 Dec 2025 GPT-5.2 Codex system card addendum 0 Irregular CyScenarioBench, text says no challenges solved (5.1.2.4) Cyber risk

More in Cybersecurity

All 111 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%
  • CWE-bench In use Agents audit a repository and write a patch for a security flaw, without being told where the bug is. Covers six programming languages. Gemini 4 Carbon 71%
  • Real-world vulnerability discovery In use Agents search popular open-source codebases to rediscover about 1,200 recent confirmed vulnerabilities across many languages. Reports share discovered. Gemini 4 Argon 86%
  • Artificial Analysis Cyber Index In use Agents do enterprise cyber defense tasks from three benchmarks, CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA, combined into one success rate. Mistral Large 4 (Le Chonk, open weights) 51%
  • Big Sleep Evaluation (Google) In use Agents search large real codebases such as Chrome and Safari for critical, hard to find vulnerabilities. Scored by pass@1 success rate, run without safety guardrails. Gemini 3.5 Flash Cyber 72%