Claude Agents Model Routing
Which AI model I point each kind of agent task at — and why.
Running AI agents well isn't just which model — it's using the right model for each task. A premium model on every job is wasteful; a cheap one on the wrong job is unreliable. So I route work deliberately: read-heavy, low-stakes drafting goes to inexpensive open models, while precise edits and anything that ships stay on a premium model where a mistake is expensive. Claude Code lets each recurring skill pin its own model in its config, so a task always runs at the cost/quality point it actually needs — independent of whatever model the interactive session happens to be on. This page is the routing map I work to.
The Routing Principle
Three questions decide the model for any task.
2. Is it read/draft, or precise edit? Reading code and drafting text tolerate a lighter model; surgical structural edits do not.
3. How much context must it hold at once? Whole-codebase or long-horizon work needs a large-context model even if the per-token price is higher.
The Model Roster
The models I route between, from premium-reliable to low-cost open.
Task → Model Map
What each kind of task currently runs on, and the reason behind the choice.
| Task | Model | Tier | Why this model |
|---|---|---|---|
Draft documentation from a codebase/update_crm_docs |
Qwen3-Coder | cheap | Reads a whole app in one large context and writes structured prose at a fraction of premium cost. Drafting tolerates a lighter model — the output is reviewed before it lands. |
Database schema & ER diagramsMermaid / DBML |
Qwen3-Coder | cheap | Diagram syntax is essentially code, so a coding model infers table relationships and cardinality accurately — and it's cheap enough to iterate on the layout. |
| Read & explain unfamiliar code | Qwen3-Coder | cheap | Read-only and long-context — no risk from a lighter model, and the large window means fewer round-trips. |
Apply drafted docs into real files/c-update_crm_doc |
Claude | premium | Precise structural edits — HTML sections, page code — where a weaker model corrupts tag nesting. Here reliability matters more than cost, so the cheap draft is applied by the premium model. |
| Production code, report/email changes — anything that ships | Claude | premium | Highest-stakes work; a mistake is expensive and hard to reverse, so it runs on the most reliable model available. |
| General reasoning when the task isn't code-shaped | GPT-OSS-120B | cheap · reliable | A dependable open model for analysis and prose, and the fallback when the coding model's endpoint is momentarily busy. |
| Quick lookups & light formatting | Llama 3.3 70B | cheap | Fast and inexpensive for low-stakes turns where a large model would be overkill. |
| Long-horizon refactoring across a large codebase | GLM-5.2 | long-context | Holds architecture, module boundaries and API contracts across a very large (~1M-token) context — worth the higher per-token price when the task spans the whole system. |
The cheap open models run through a single low-cost gateway on a separate Claude Code profile; the premium tier is the standard Claude account. A skill can hard-pin any of these in its config so the choice is made once, not re-decided every run.