Benchmarks / Cyber
SRE Bench
Agent reverse engineering of real-world binary programs without access to source code.
SRE Bench
Higher is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | GPT-6 Astra | max | 56.9% |
| 2 | GPT-6.1 Sol | max | 50.8% |
| 3 | Gemini 4 | high | 44.3% |
| 4 | Opus 5.5 | max | 33.6% |
| 5 | Daybreak Blue | max | 31.7% |
| 6 | GPT-5.6 Sol | max | 30.5% |
| 7 | GPT-6 Sol | max | 30.5% |
| 8 | Sonnet 5.5 | max | 30.2% |
| 9 | Fable 5.1 | max | 22.9% |
| 10 | Opus 5 | max | 12.2% |
| 11 | Wonder Full | max | 11.8% |
| 12 | Gemini 3.8 Flash | high | 11.1% |
| 13 | GPT-5.6 Terra | max | 9.5% |
| 14 | GLM 5.3 | max | 9.2% |
| 15 | Haiku 5.5 | max | 5.7% |
| 16 | Gemini 3.7 Flash | high | 4.6% |
| 17 | MiMo V2.6 Flash | high | 4.2% |
| 18 | Grok 4.7 | xhigh | 3.8% |
| 19 | GPT-5.5 | xhigh | 3.8% |
| 20 | Muse Spark 1.3 | max | 3.8% |
| 21 | MiMo V2.6 Pro | high | 3.1% |
| 22 | GPT-6 Luna | max | 2.7% |
| 23 | Hy4 | – | 2.3% |
| 24 | DeepSeek V4 Pro 0813 | max | 1.9% |
| 25 | GPT-5.6 Luna | max | 1.5% |
| 26 | DeepSeek V4.1 Flash | max | 0.8% |
| 27 | GLM 5.3 Flash | max | 0.8% |
| 28 | Grok 4.5 | high | 0.8% |
| 29 | Sonnet 5 | max | 0.4% |
| 30 | Step 5 Preview | – | 0.4% |
| 31 | GLM 5.2 | max | 0.0% |
| 32 | Muse Spark 1.2 | xhigh | 0.0% |
| 33 | Gemini 3.1 Pro | high | 0.0% |
| 34 | Inkling | max | 0.0% |