Downgraded Claude Max 20x to 5x after moving the boring half of my work to my laptop
retained text
I was paying $200 a month for 20x and when I actually looked at what I was burning it on, a lot of it was dumb stuff. Summarizing long threads, explaining old files before I touch them, commit messages, renaming things. Opus doing jobs a small model could do. So I started running a small model on my base M4 16GB next to Claude Code. Gemma 4 E4B through Magnitude, it reads a 2k token prompt and starts answering in about 2 seconds. That wait was what made local feel useless before, Ollama took over 6 seconds every time on the same model. It's not doing real work though. I tried plugging it in as the actual Claude Code model and a 4B model just falls over with that many tools. As a side chat for the read and explain stuff it's fine, and my usage dropped enough that 5x has been fine for 3 weeks. That's $100 a month back. Anyone else splitting work like this or is everyone just going all in on one plan?
document fetched 2026-11-14 00:45 · canonical https://reddit.com/r/ClaudeCode/comments/1xv6lxa/downgraded_claude_max_20x_to_5x_after_moving_the