Claude Repeats Same Random Word

This looks more like a sampling expectation mismatch than a bug: if the prompt, model snapshot, and hidden system context stay the same, Claude can keep landing on the same high-probability token.

• –Anthropic’s docs say `temperature` controls randomness, but even `temperature: 0.0` is not fully deterministic
• –Claude’s defaults lean toward creative output, yet a tiny prompt can still collapse onto a favorite token like “ephemeral”
• –Starting a new chat does not necessarily mean a new seed, a new model version, or a meaningfully different context
• –The fact that other users got different words suggests subtle differences in model state, settings, or account/session context
• –If you want actual variety, force a choice set or ask for multiple candidates instead of a single “random” word