I Replaced My $200/Month Claude Code with These Local Models.

One model for the agent, one for coding, one for autocomplete, and what fits on 24GB, 48GB and 64GB Macs.

Read the article for free here.

I moved from Claude’s $200 Max plan to the $20 Pro plan. Pro still includes Claude Code, with a twentieth of the usage, so the plan is to save it for the hard tasks and send the routine work somewhere else. Mine goes to three open models running on a 48GB M5 Pro.

Coding with an assistant is three jobs. An agent like Claude Code sends 16k to 20k tokens of instructions and tool descriptions before you’ve typed anything, then works through a task over many turns. Normal coding is a question or a single function, with a short prompt and an answer you want to be right. Autocomplete suggests the next few lines as you type and has to keep up with your typing. No single open model does all three well. The agent job needs a model that reads prompts fast, the coding job needs the best answers you can get, and autocomplete needs a small model trained to fill in code.

So the stack uses one model for each job. You switch between the agent model and the coding model as the work changes, and the autocomplete model runs alongside whichever one is active. How much of this runs…