Skip to main content

Cost & model routing

Not every role needs your most expensive model. lmctl lets you route models by role: put top-tier models where judgment pays off, and cheaper-but-capable models on high-volume mechanical work.

  • Top-tier models — the Lead (planning, delegation, whole-project picture), the designer, research, and the reviewer. These roles make decisions; pay for quality there.
  • Leaner, cheaper models — a Coder doing heavy read/write on a well-scoped task. The work is high volume and mechanical; a capable mid-tier model handles it well at a fraction of the cost.

This pairs directly with division of context: a well-scoped Coder doesn't need a frontier model or a giant context window.

Role
What it decides
Model posture
Why
Lead
Plan, split work, arbitrate reviews
Top-tier
Judgment errors propagate across the team.
Coder
Implement scoped changes
Lean capable model
High-volume work benefits from lower cost per turn.
Reviewer
Challenge the result
Strong independent model
Review quality depends on seeing what the author missed.

Routing models in the teamfile​

Set the model per member with model= on each _MEMBER_ line. Use effort=<variant> to pick a reasoning variant for OpenCode reasoning models.

_MEMBER_ alias=Lead     provider=claude    model=<top-tier-id>
_MEMBER_ alias=Coder    provider=qwen      model=qwen3-coder-plus
_MEMBER_ alias=Reviewer provider=codex     model=<top-tier-id>

A top-tier Lead and Reviewer keep the judgment sharp; a leaner Coder does the bulk of the typing. lmctl lint <teamfile.lmctl> validates the models you pick against the tested catalog. effort= selects a provider variant such as an OpenCode reasoning-effort profile; lmctl rejects or warns on unsupported provider/effort combinations.

Version floor: use @lmctl-ai/lmctl 0.1.151 or newer for model-routed teams (verified against 0.1.218). Earlier public-preview builds could silently ignore model= during some seed or terminal paths. After seeding, run lmctl health <teamfile.lmctl> and confirm the MODEL column matches the teamfile before trusting a routed run.

Cheap-and-capable coder models​

Through the OpenCode provider, the Coder seat opens up to cost-effective models — including some that are essentially free to run:

  • DeepSeek — strong coding quality at low cost.
  • Qwen-coder — capable, inexpensive coding model.
  • GPT-5-mini and other small hosted models — cheap per-token, fine for scoped tasks.
  • Local models via Ollama — free to run on your own hardware.

Wire these up in an opencode.json — see the sample config for current provider examples.

We're deliberately not quoting prices here: provider pricing changes often, and a number that's right today is wrong next quarter. The durable advice is the shape — match model cost to role, and lean on OpenCode for the cheap seats.