Free Claude Code: Match Every Task to the Right AI Model

Your cheapest tasks are running on your most expensive model. Not because you chose that, but because nobody chose anything — the default quietly routes everything through the same brain, including the work that needed almost none.

A free repo with 45,000 stars fixes that by splitting your setup in two: the harness you built, and the brain you rent. Once those come apart, you can point each kind of work at the model that should have been doing it all along.

Here is the full breakdown, concept by concept.

The short version

• Harness = what you built. Brain = what you rent. Alignment = what you keep.
• By default a subagent runs on whatever model the parent was using, so chores bill at heavy-reasoning rates.
• Each model tier can point at a different provider — including free and local ones.

1. The harness and the brain

The single most useful idea here is the separation itself.

The harness is local. Your prompts, your skills, your tools, your project conventions, the accumulated setup that makes the thing useful for your work specifically. That is the part you spent months building and the part that carries your value.

The brain is the model underneath. Rented, interchangeable, and — once the two are properly separated — swappable without touching anything else.

Most people treat these as one product. They are not, and once you stop treating them as one, the model becomes a runtime choice rather than an identity.

2. Why the default costs you

Here is the specific leak, and it is not obvious from the outside.

By default, a subagent runs on whatever model the parent was using. You start a session on a heavy model because the architectural question at the top was genuinely hard. Then the session spawns subagents to do the follow-on work — and they inherit the same model.

So a one-line file rename bills at heavy-reasoning rates. So does a lookup, a formatting pass, a classification, and every other mechanical step in a long session. Individually they are trivial. In aggregate, across a working week, they are most of the invoice.

The parent model choice was correct. The inheritance is what costs you.

3. Model-task alignment

The fix is to match the brain to the job, tier by tier:

• Opus for the hard architectural thinking — the decisions you will build on top of for months.
• Sonnet for daily writing and coding, which is the bulk of the work.
• Haiku or a local model for the chores nobody should be paying premium rates for.

This is not a compromise on quality. A file rename does not get better on a larger model; it gets more expensive. Alignment means spending capability where capability changes the outcome, and nowhere else.

4. Bring any brain

Once the harness is separated, the set of available brains gets much larger than the set your subscription includes.

Connect free, paid or local models. OpenRouter alone reaches hundreds of models behind a single API, which turns model choice into a configuration line rather than a migration.

And for the bottom tier, you may not need a provider at all: Ollama and LM Studio run models on your own machine, for nothing. A local model handling the chores costs electricity, and the chores are precisely the tasks where a local model is good enough.