Repository navigation
Model Tiers
Definition. A tier is the portable label an agent declares for how much model it needs:
fast, balanced or strong. anycrew maps the tier to each assistant's own model vocabulary, so
the agent files stay the same everywhere and the cheap work goes to cheap models.
| Tier | Purpose | Cost of being wrong | Agents |
|---|---|---|---|
| fast | read, search, summarise, format, status updates | a minute | scout, doc-writer |
| balanced | implement a task, write tests, research, infra | an hour | implementer, tester, researcher, designer, release-engineer |
| strong | decisions and judgement: scope, architecture, specs, security, review, hard bugs | a week | brainstormer, product-manager, architect, spec-writer, reviewer, security-reviewer, debugger, ai-engineer, crew-lead |
models.json in the repository maps each tier per assistant: Claude Code aliases plus effort
(haiku/low, sonnet/medium, opus/high), Antigravity tiers (flash, flash, pro), Codex
reasoning effort (low, medium, high), OpenCode provider/model, and inherit for Cursor and
Copilot until you set model IDs. Projects override per tool in .anycrew/config.json under
models; anycrew list shows the effective table.
Model names are tool-specific and change monthly. A tier is a statement about the job, which does not change. The mapping is one small file to maintain instead of sixteen agents times six tools.
Start one tier lower than instinct says and promote when reviews or evals show the agent falling short. The cheapest model that passes is the right one.