CWE-bench
Agents audit a repository and write a patch for a security flaw, without being told where the bug is. Covers six programming languages.
V1 has 120 tasks and is not comparable with v0. The held-out set is private to Collinear. Google reports Gemini 4 Argon at 68% pass@1 and says patches must preserve the existing functional and regression tests.
- Made by
- Collinear AI
- Released
- September 2026
- Kind
- Secure code and review
- Level
- Expert
- Scored
- Checked automatically
- Answers
- Code
- Tasks
- Kept private
- Strongest model tested
-
Gemini 4 Carbon #2 on the capability index - Last in a launch post
- October 2026
- Last model added
- October 2026
- Scores refreshed
- November 12th, 2026
Gemini 4 Carbon leads with 71%
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.
Gemini 4 Carbon leads with 71%, claims included
Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.
- 1 Gemini 4 Carbon 71%
- 2 GPT-6 Astra 68%
- 3 Grok 4.7 68%
- 4 Gemini 4 Argon 68%
- 5 Claude Opus 5.5 67%
- 6 Claude Fable 5.1 58%
- 7 Grok 4.6 57%
- 8
Muse Spark 1.3 55% - 9
DeepSeek V4.1 Flash 55% - 10
Tencent Hy4 preview 53% - 11 GPT-6 Sol 52%
- 12
Inkling 37%
Source: Official CWE-bench leaderboard
- 1 Gemini 4 Carbon 71%
- 2 GPT-6 Astra 68%
- 3 Grok 4.7 68%
- 4 Gemini 4 Argon 68%
- 5 Claude Opus 5.5 67%
- 6 Claude Fable 5.1 58%
- 7 Grok 4.6 57%
- 8 Muse Spark 1.3 55%
- 9 DeepSeek V4.1 Flash 55%
- 10 Tencent Hy4 preview 53%
- 11 GPT-6 Sol 52%
- 12 Gemini 3.8 Flash Cyber 47.2%
- 13 Inkling 37%
Sources: Official CWE-bench leaderboard and labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Gemini 4 Carbon | 71% | Antigravity | Oct 2026 | Official leaderboard |
| 2 | GPT-6 Astra | 68% | Codex | Sep 2026 | Official leaderboard |
| 3 | Grok 4.7 | 68% | opencode | Sep 2026 | Official leaderboard |
| 4 | Gemini 4 Argon | 68% | Antigravity | Sep 2026 | Official leaderboard |
| 5 | Claude Opus 5.5 | 67% | Claude Code | Sep 2026 | Official leaderboard |
| 6 | Claude Fable 5.1 | 58% | Claude Code | Sep 2026 | Official leaderboard |
| 7 | Grok 4.6 | 57% | opencode | Aug 2026 | Official leaderboard |
| 8 | Muse Spark 1.3 | 55% | opencode | Sep 2026 | Official leaderboard |
| 9 | DeepSeek V4.1 Flash | 55% | opencode | Sep 2026 | Official leaderboard |
| 10 | Tencent Hy4 preview | 53% | opencode | Aug 2026 | Official leaderboard |
| 11 | GPT-6 Sol | 52% | Codex | Sep 2026 | Official leaderboard |
| 12 | Inkling | 37% | opencode | Jul 2026 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Gemini 4 Carbon | 71% | Antigravity | Oct 2026 | Official leaderboard |
| 2 | GPT-6 Astra | 68% | Codex | Sep 2026 | Official leaderboard |
| 3 | Grok 4.7 | 68% | opencode | Sep 2026 | Official leaderboard |
| 4 | Gemini 4 Argon | 68% | Antigravity | Sep 2026 | Official leaderboard |
| 5 | Claude Opus 5.5 | 67% | Claude Code | Sep 2026 | Official leaderboard |
| 6 | Claude Fable 5.1 | 58% | Claude Code | Sep 2026 | Official leaderboard |
| 7 | Grok 4.6 | 57% | opencode | Aug 2026 | Official leaderboard |
| 8 | Muse Spark 1.3 | 55% | opencode | Sep 2026 | Official leaderboard |
| 9 | DeepSeek V4.1 Flash | 55% | opencode | Sep 2026 | Official leaderboard |
| 10 | Tencent Hy4 preview | 53% | opencode | Aug 2026 | Official leaderboard |
| 11 | GPT-6 Sol | 52% | Codex | Sep 2026 | Official leaderboard |
| 12 | Gemini 3.8 Flash Cyber | 47.2% | Sep 2026 | Lab model card | |
| 13 | Inkling | 37% | opencode | Jul 2026 | Official leaderboard |
Labs report scores for 3 models in their cards. Switch the scores above to rank them together.
The best score, 71%, costs $5.83 a task
Runs on the cost frontier score higher than any run that's cheaper. Claude Opus 5.5 comes within 4 points of the best for $0.79 a task, 7 times cheaper. The costs are the leaderboard's own estimates, so compare them only within this benchmark.
Source: Official CWE-bench leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Gemini 4 Carbon | Antigravity | 71% | $5.83 | ||
| GPT-6 Astra | Codex | 68% | $2.85 | ||
| Grok 4.7 | opencode | 68% | $2.75 | ||
| Gemini 4 Argon | Antigravity | 68% | $6.63 | ||
| Claude Opus 5.5 | Claude Code | 67% | $0.79 | ||
| Claude Fable 5.1 | Claude Code | 58% | $3.47 | ||
| Grok 4.6 | opencode | 57% | $1.09 | ||
| Muse Spark 1.3 | opencode | 55% | $3.95 | ||
| DeepSeek V4.1 Flash | opencode | 55% | $0.09 | ||
| Tencent Hy4 preview | opencode | 53% | $0.47 | ||
| GPT-6 Sol | Codex | 52% | $0.96 | ||
| Inkling | opencode | 37% | $3.00 |
3 models have held number one on the official leaderboard since its release in September 2026 3 models have held number one with claims included since its release in September 2026
Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.
Sources: Official CWE-bench leaderboard and labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Sources: Official CWE-bench leaderboard and labs’ model cards
One lab put it in a launch post in the last six months
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| Oct 28, 2026 | Gemini 4 Carbon | 71 | CWE-bench |
| Oct 13, 2026 | Gemini 4 Argon | 68 | CWE-bench v1 (Cybersecurity row of headline table) |
| Sep 2, 2026 | Gemini 3.8 Flash Cyber | 47.2 | CWE-Bench (run by Collinear), pass@1 |
2 system cards report it, from 1 lab
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.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Oct 13, 2026 | Gemini 4 Argon system card | 68 | CWE-Bench v1 (120 tasks), Pass@1 mean of 4 trials at high reasoning 68% (7.2) | Cyber risk |
| Oct 13, 2026 | Gemini 4 Argon model card | 68 | CWE-bench v1, Cybersecurity row (results table, p.5) | Cyber risk |
2 lab claims next to the leaderboard
Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. 2 of them match, which usually means one side copied the other. Every benchmark.
| Model | Lab says | Board says | Gap |
|---|---|---|---|
| Gemini 4 Argon | 68% CWE-bench v1 (Cybersecurity row of headline table) | 68% Leaderboard, Antigravity | Same |
| Gemini 4 Carbon | 71% CWE-bench | 71% Leaderboard, Antigravity | Same |
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.
Source: Official leaderboards
More 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%
- CyberBench v1.1 (Vals) In use 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. Muse Wonder 80.4%
- 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%