LLMs are super biased coin flippers

if you ask claude opus 4.6 to “flip a coin” 100,000 times via anthropic’s batch API, it will

• say “heads” 94127 times
• be ambigous 5873 times
• say “tails” 0 times