Model Tiers

Model tiers provide semantic model selection for portable workflows.

Overview

Instead of hardcoding model names, use tiers that describe capability needs:

steps:
  - id: classify
    type: llm
    model: fast
    prompt: Classify this as bug, feature, or question

  - id: review
    type: llm
    prompt: Review this code for issues

  - id: architect
    type: llm
    model: strategic
    prompt: Design a scalable architecture

The Three Tiers

Fast

Low-latency, cost-effective models for simple tasks.

Use for:

Text classification
Simple extraction
Formatting and summarization
High-volume automation

Characteristics:

Latency: <500ms typical
Cost: Lowest per token
Capability: Basic reasoning

Balanced

General-purpose models for most workflows. This is the default if you omit model.

Use for:

Code review
Content generation
Analysis and synthesis
Multi-step reasoning

Characteristics:

Latency: <2s typical
Cost: Moderate
Capability: Strong reasoning

Strategic

Most capable models for complex reasoning.

Use for:

Complex analysis
Research synthesis
Creative problem-solving
Multi-domain expertise

Characteristics:

Latency: <10s typical
Cost: Highest per token
Capability: Best available reasoning

Provider Mappings

Tier	Anthropic	OpenAI	Google
fast	Claude 3 Haiku	GPT-3.5 Turbo	Gemini Flash
balanced	Claude 3.5 Sonnet	GPT-4 Turbo	Gemini Pro
strategic	Claude Opus 4.5	GPT-4	Gemini Ultra

Choosing a Tier

Is the task simple pattern matching?
├── Yes → fast
└── No → Does it require complex reasoning?
    ├── Yes → Does it involve novel problem-solving?
    │   ├── Yes → strategic
    │   └── No → balanced
    └── No → balanced

Tip: Start with balanced (the default). Downgrade to fast if results are good, upgrade to strategic if needed.

Multi-Tier Workflows

Use different tiers for different steps:

steps:
  - id: review
    type: llm
    prompt: Review this code for issues

  - id: format
    type: llm
    model: fast
    prompt: "Format this review as markdown: {{.steps.review.response}}"