Skip to content

Model Tiers

Ankit Jha edited this page May 29, 2026 · 1 revision

Model Tiers

Hydra routes tasks to one of ten tiers. The tier is selected by resolving a complexity enum key (e.g. SIMPLE) to a number defined in registry/routing.yaml.


Full Tier Reference

Tier Enum Model Executor Role Cost
1 CORE Claude Code Direct CLI Brain — orchestrates, never dispatched $$$$$
2 EXPERT Claude Opus 4.6 (Thinking) Antigravity Architecture decisions, distributed systems $$$$
3 VERY_HARD Claude Sonnet 4.6 (Thinking) Antigravity Complex refactors, multi-step reasoning $$$
4 HARD Gemini 2.0 Flash Thinking Antigravity Tradeoff reviews, rubber duck for Claude output $$
5 COMPLEX Gemini 3.1 Pro (High) Antigravity Middleware, auth integrations, ML pipelines $$
6 MODERATE Gemini 3.1 Pro (Low) Antigravity Services, repositories, validators $
7 STANDARD Gemini 3.5 Flash (High) Antigravity Controllers, handlers, basic CRUD $
8 SIMPLE Gemini 3.5 Flash (Medium) Antigravity DTOs, interfaces, database schemas $
9 TRIVIAL Gemini 3.5 Flash (Low) Antigravity Constants, enums, simple type updates $
10 GRUNT Qwen2.5-Coder 7B Local Ollama Boilerplate, mock data, simple scripts Free

Tier Groups

Expert (T2–T3): Claude Family

Use for tasks that require deep reasoning, multi-file analysis, or architectural judgment.

  • T2 EXPERT — Genuine architecture/research work only. Expensive, shares pool with Sonnet.
  • T3 VERY_HARD — Preferred Claude tier for complex refactors. Cheaper than Opus, still thinking-mode.

Review (T4): Gemini Thinking

T4 HARD uses Gemini 2.0 Flash Thinking — a fast thinking-mode model used as a rubber duck reviewer for Claude output. Being a different model family helps surface blind spots that Claude-family models share.

Standard (T5–T8): Gemini Family

The workhorse tiers. Most developer tasks land here.

  • T5–T6: Gemini Pro — accurate enough for middleware and service logic.
  • T7–T8: Gemini Flash — fast and cheap for controllers and schemas.
  • T9: Flash Low — borderline accuracy, trivial config changes only.

Grunt (T10): Local Ollama

T10 GRUNT runs Qwen2.5-Coder 7B locally. Free, always available, no network dependency. The terminal fallback in every chain.


Token Pools

Pool Members Notes
agy_claude T2, T3 Shared Opus + Sonnet allocation
agy_gemini_pro T5, T6 Shared Pro High + Pro Low
agy_flash T4, T7, T8, T9 High-volume Flash allocation
local_ollama T10 Unlimited local — no quota

Clone this wiki locally