KillSwitch-Bench pits 20 coding agents against an LLM-hostile language

KillSwitch-Bench published a leaderboard of 20 model-plus-harness combinations writing code in KillSwitch, a stack language designed to be hard for LLMs to read or write. Claude Opus 5 leads at 66.8%, while GPT-5.6 Luna and Qwen3.8 27B scored 0.0% at $0.00 per task.

• Claude Opus 5 tops the board at 66.8%, followed by GPT-6.1 Sol at 59.2%
• GPT-5.6 Luna and Qwen3.8 27B scored 0.0% at zero cost per task
• Claude Sonnet 5.5 is 13x faster and 17x cheaper than Sonnet 5 but scores half as much
• The AI_SPEC.md file is roughly 1.3M tokens and exceeds any context window