Cyber Index
Artificial Analysis ranking of models on cybersecurity tasks: successes, safety blocks and cost per task.
- Configurations
- 20
- Best score
- 68.7%
- GPT-6 Sol (Daybreak Blue, max) · Trusted access
- Best public model
- 56.4%
- Grok 4.7 (Xhigh)
- With safety blocks
- 7
- of 20 configurations
Score vs cost per task
Each point is a measured configuration. Top right: the best score for the lowest cost. Hover or move through the points with the keyboard to read their values.
Models shown20 of 20
Select a model to show or hide it; hovering it brings it forward on the chart. The axes fit the models shown.
20 / 20 configurations
- Successes
- Safety blocks
- Failures
- Trusted access (not public)
- #1GPT-6 Sol (Daybreak Blue, max)Trusted access. A version reserved for a trusted-access programme, not publicly available, evaluated without the standard safeguards of the public model.68.7%
OpenAI
68.7% successes, 0.0% safety blocks
- CWE-Bench-AA
- 64.7%
- DeepsecBench-AA
- 48.1%
- CyberGym-E2E-AA
- 93.1%
- Cost / task
- $1.77
56.4% successes, 0.0% safety blocks
- CWE-Bench-AA
- 68.3%
- DeepsecBench-AA
- 26.9%
- CyberGym-E2E-AA
- 74.0%
- Cost / task
- $11.67
56.1% successes, 0.0% safety blocks
- CWE-Bench-AA
- 63.3%
- DeepsecBench-AA
- 26.2%
- CyberGym-E2E-AA
- 78.6%
- Cost / task
- $0.184
52.7% successes, 0.0% safety blocks
- CWE-Bench-AA
- 56.7%
- DeepsecBench-AA
- 23.5%
- CyberGym-E2E-AA
- 77.9%
- Cost / task
- $0.117
49.9% successes, 0.0% safety blocks
- CWE-Bench-AA
- 55.8%
- DeepsecBench-AA
- 19.9%
- CyberGym-E2E-AA
- 74.0%
- Cost / task
- $0.356
49.5% successes, 0.0% safety blocks
- CWE-Bench-AA
- 50.8%
- DeepsecBench-AA
- 16.0%
- CyberGym-E2E-AA
- 81.7%
- Cost / task
- $2.79
47.7% successes, 0.0% safety blocks
- CWE-Bench-AA
- 52.5%
- DeepsecBench-AA
- 15.8%
- CyberGym-E2E-AA
- 74.8%
- Cost / task
- $1.80
44.3% successes, 0.0% safety blocks
- CWE-Bench-AA
- 60.8%
- DeepsecBench-AA
- 12.5%
- CyberGym-E2E-AA
- 59.5%
- Cost / task
- $3.07
41.3% successes, 0.0% safety blocks
- CWE-Bench-AA
- 53.3%
- DeepsecBench-AA
- 12.5%
- CyberGym-E2E-AA
- 58.0%
- Cost / task
- $2.45
40.8% successes, 0.0% safety blocks
- CWE-Bench-AA
- 68.3%
- DeepsecBench-AA
- 31.2%
- CyberGym-E2E-AA
- 22.9%
- Cost / task
- $0.486
36.8% successes, 36.4% safety blocks
- CWE-Bench-AA
- 64.2% (safety blocks)
- DeepsecBench-AA
- 46.3%
- CyberGym-E2E-AA
- 0.0% (safety blocks)
- Cost / task
- $1.96
36.4% successes, 0.0% safety blocks
- CWE-Bench-AA
- 55.8%
- DeepsecBench-AA
- 24.3%
- CyberGym-E2E-AA
- 29.0%
- Cost / task
- $2.43
33.4% successes, 37.8% safety blocks
- CWE-Bench-AA
- 63.3% (safety blocks)
- DeepsecBench-AA
- 36.9%
- CyberGym-E2E-AA
- 0.0% (safety blocks)
- Cost / task
- $4.98
28.7% successes, 35.9% safety blocks
- CWE-Bench-AA
- 58.3% (safety blocks)
- DeepsecBench-AA
- 26.9%
- CyberGym-E2E-AA
- 0.8% (safety blocks)
- Cost / task
- $9.75
25.1% successes, 31.8% safety blocks
- CWE-Bench-AA
- 52.5% (safety blocks)
- DeepsecBench-AA
- 14.5%
- CyberGym-E2E-AA
- 8.4% (safety blocks)
- Cost / task
- $3.22
25.0% successes, 38.4% safety blocks
- CWE-Bench-AA
- 50.8% (safety blocks)
- DeepsecBench-AA
- 23.5%
- CyberGym-E2E-AA
- 0.8% (safety blocks)
- Cost / task
- $8.09
22.4% successes, 0.0% safety blocks
- CWE-Bench-AA
- 41.7%
- DeepsecBench-AA
- 1.1%
- CyberGym-E2E-AA
- 24.4%
- Cost / task
- $1.08
13.1% successes, 0.0% safety blocks
- CWE-Bench-AA
- 27.5%
- DeepsecBench-AA
- 1.0%
- CyberGym-E2E-AA
- 10.7%
- Cost / task
- $6.67
12.7% successes, 56.8% safety blocks
- CWE-Bench-AA
- 26.7% (safety blocks)
- DeepsecBench-AA
- 11.5% (safety blocks)
- CyberGym-E2E-AA
- 0.0% (safety blocks)
- Cost / task
- $1.54
12.6% successes, 62.5% safety blocks
- CWE-Bench-AA
- 28.3% (safety blocks)
- DeepsecBench-AA
- 9.5% (safety blocks)
- CyberGym-E2E-AA
- 0.0% (safety blocks)
- Cost / task
- $2.61
| # | Model | ||||||
|---|---|---|---|---|---|---|---|
GPT-6 Sol (Daybreak Blue, max)Trusted access. A version reserved for a trusted-access programme, not publicly available, evaluated without the standard safeguards of the public model. OpenAI | 68.7% | 0.0% | 64.7% | 48.1% | 93.1% | $1.77 | |
| 2 | Grok 4.7 (Xhigh) SpaceXSI | 56.4% | 0.0% | 68.3% | 26.9% | 74.0% | $11.67 |
| 3 | MiMo-V2.6-Pro Xiaomi | 56.1% | 0.0% | 63.3% | 26.2% | 78.6% | $0.184 |
| 4 | GPT-6 Luna (Max) OpenAI | 52.7% | 0.0% | 56.7% | 23.5% | 77.9% | $0.117 |
| 5 | GLM-5.3-Flash Z.AI | 49.9% | 0.0% | 55.8% | 19.9% | 74.0% | $0.356 |
| 6 | Mistral Large 4 Preview Mistral | 49.5% | 0.0% | 50.8% | 16.0% | 81.7% | $2.79 |
| 7 | Step 5 Preview StepFun | 47.7% | 0.0% | 52.5% | 15.8% | 74.8% | $1.80 |
| 8 | 44.3% | 0.0% | 60.8% | 12.5% | 59.5% | $3.07 | |
| 9 | Kimi K3 (Max) Moonshot AI | 41.3% | 0.0% | 53.3% | 12.5% | 58.0% | $2.45 |
| 10 | DeepSeek V4.1 Flash (Max) DeepSeek | 40.8% | 0.0% | 68.3% | 31.2% | 22.9% | $0.486 |
| 11 | GPT-6 Sol (Max) OpenAI | 36.8% | 36.4% | 64.2% (safety blocks) | 46.3% | 0.0% (safety blocks) | $1.96 |
| 12 | GLM-5.3 (Max) Z.AI | 36.4% | 0.0% | 55.8% | 24.3% | 29.0% | $2.43 |
| 13 | GPT-6 Astra (Max) OpenAI | 33.4% | 37.8% | 63.3% (safety blocks) | 36.9% | 0.0% (safety blocks) | $4.98 |
| 14 | 28.7% | 35.9% | 58.3% (safety blocks) | 26.9% | 0.8% (safety blocks) | $9.75 | |
| 15 | Gemini 3.8 Flash (High) | 25.1% | 31.8% | 52.5% (safety blocks) | 14.5% | 8.4% (safety blocks) | $3.22 |
| 16 | 25.0% | 38.4% | 50.8% (safety blocks) | 23.5% | 0.8% (safety blocks) | $8.09 | |
| 17 | MiniMax-M3 MiniMax | 22.4% | 0.0% | 41.7% | 1.1% | 24.4% | $1.08 |
| 18 | 13.1% | 0.0% | 27.5% | 1.0% | 10.7% | $6.67 | |
| 19 | Qwen3.8 27B (Xhigh) Alibaba | 12.7% | 56.8% | 26.7% (safety blocks) | 11.5% (safety blocks) | 0.0% (safety blocks) | $1.54 |
| 20 | Qwen3.8 2.4T A95B Alibaba | 12.6% | 62.5% | 28.3% (safety blocks) | 9.5% (safety blocks) | 0.0% (safety blocks) | $2.61 |
* The model or provider declined some tasks on safety grounds.
Method and versions
The index combines the evaluations below, measured independently by Artificial Analysis. Each row is one measured configuration; its name gives the reasoning level used.
Safety blocks: Share of tasks the model or its provider declined on safety grounds. These tasks do not count as solved.
Trusted access (not public): A version reserved for a trusted-access programme, not publicly available, evaluated without the standard safeguards of the public model.
Data from Artificial Analysis, updated with each catalogue refresh. Ranks follow the index after filtering.