Atlas

Benchmarks

← All benchmarks

SRE Bench — Fully Solved

Code · 2026-09-21

Reverse engineering 262 protected binary instances built from 19 in-house programs. Agents use one shell with Ghidra, radare2, GDB, angr and other analysis tools. An instance counts as fully solved only if all six scenario-specific subtasks pass deterministic verification.

Top models (higher is better)

ModelScore
GPT-6 Astra56.9
GPT-6.1 Sol50.8
Gemini 4 Argon44.3
GPT-5.6 Sol30.5
GPT-6 Sol30.5
Muse Wonder Full11.8
Gemini 3.8 Flash11.1
GPT-5.6 Terra9.5
GLM-5.39.2
Claude Haiku 5.55.7
Gemini 3.7 Flash4.6
MiMo-V2.6-Flash4.2
GPT-5.53.8
Grok 4.73.8
Muse Spark 1.33.8
MiMo-V2.6-Pro3.1
GPT-6 Luna2.7
Hy4 Preview2.3
DeepSeek V4 Pro 08131.9
GPT-5.6 Luna1.5
DeepSeek V4.1 Flash0.8
GLM-5.3 Flash0.8
Grok 4.50.8
Sonnet 50.4
Step 5 Preview0.4
Gemini 3.1 Pro Preview0.0
GLM-5.20.0
Inkling0.0
Muse Spark 1.20.0
Loading Atlas data…