I'm currently using Claude code through my Max subscription, and I'm starting to wonder if Anthropic is utilizing models other than Sonnet. Recently, I encountered a bizarre situation where Claude tried to generate a random API key instead of recognizing the real one I have in my environment variables. When I inquired about this, Claude insisted it found the key in my code, which definitely wasn't true since the imaginary key didn't conform to standard key formats. This leads me to think that it might be a smaller model; after all, Sonnet should know that you can't just invent an API key out of thin air, right?
4 Answers
I think I can help with that! Claude does store some conversation metadata in a local file path, specifically in ~/.claude/local/local/__store.db. If you dig in there, you can query for metadata related to token and model usage through SQLite. Just keep that in mind while troubleshooting!
I ran into a similar experience. A few days ago, mine switched to using 3.5 Sonnet after I cleared the cache, and that seemed to fix a lot of quirks! Just something to consider.
Claude definitely employs Haiku as well; it's not just Sonnet. However, don't forget that all models, including Sonnet, can have their off days and may sometimes hallucinate. It's kind of a known issue across different architectures!
Yeah, that's just how LLMs operate sometimes! Hallucinations are pretty common. One tip is to use the /clear command more often. Honestly, I'm a huge fan of Claude and its capabilities; it really does impress me!
That's really helpful, thanks!