Hacker News Reader

If you change reasoning during a session, LLM needs to recompute KV cache. Maybe I get it wrong, but I don't think it can save money.

The newer models from both OpenAI and Anthropic allow you to dynamically change reasoning effort without resetting cache