HellGates, custom CPU gate-level challenge

Automated excerpt

It was designed for humans, but apparently worked well against LLMs too, until now. For a year, nobody solved it. Not the humans, they spent weeks/months on it and gave up. Not the LLMs, Claude failed, ChatGPT failed, and DeepSeek failed after days or weeks of work guided by hints I gave to people, ending with the model declaring the challenge “computationally infeasible with available resources. ”Then, in September 2026, GPT-6 solved it in under 20-30 minutes on SRE-Bench. Thank you Zhuo Zhang for making this possible !

Selected automatically from source text; not independently written or fact-checked. Read the original for full context.