Delegation Policy
Route tasks to the right model and agent for optimal quality per token.
Model Tiers (2026)
| Tier | Anthropic | OpenAI | OpenCode Go | Cost | |
|---|---|---|---|---|---|
| Fast | Claude Haiku 4.5 | GPT-5.4-mini | Gemini 3 Flash | GLM-5, MiniMax M2.5 | 1x |
| Balanced | Claude Sonnet 4.6 | GPT-5.4 | Gemini 3.1 Pro | Kimi K2.5, MiniMax M2.7 | 3x |
| Deep | Claude Opus 4.6 | GPT-5.3 Codex | — | — | 5x |
Agent Routing
| Agent CLI | Best for | Default tier |
|---|---|---|
claude | Architecture, complex reasoning, multi-file refactor | Deep (opus) |
codex | Fast edits, scaffolding, mechanical changes | Balanced (gpt-5.4-mini) |
gemini | Large-context analysis, docs, code review | Balanced (gemini pro) |
aider | Git-native focused edits, test writing | Balanced (configurable) |
opencode | General-purpose, prototyping | Balanced (configurable) |
When to escalate
- Same bug investigated twice without root cause → Deep tier
- Architecture decision with unclear trade-offs → Deep tier
- Security implications span multiple systems → Deep tier
- Need to search code or read logs → Fast tier
- Need large codebase analysis → Gemini (large context)
- Need many small focused edits → Codex or Aider
Target distribution
70% balanced, 20% fast, 10% deep = ~60% cost of all-deep with equivalent quality.