Model Routing

Model routing automatically selects the right model tier for each prompt based on task complexity. Simple tasks go to cheaper, faster models; complex tasks go to more capable ones. When enabled, prompts are classified using keyword analysis and length heuristics. See configuration to enable and customize routing.

Overview

When routing is enabled, each prompt is classified into one of three tiers:

The provider then selects the appropriate model for the tier.

Configuration

[agent.routing]
enabled = false                # disabled by default
low_keywords = []              # additional keywords for low tier
high_keywords = []             # additional keywords for high tier

Classification algorithm

Prompts are classified using keyword analysis and length heuristics:

Built-in keywords

Low-tier keywords: typo, rename, format, lint, fix indent, fix spacing, fix whitespace, simple, trivial, minor, small, quick, one-line, single, update comment, add comment, remove comment

High-tier keywords: architect, design, refactor, security, audit, migrate, complex, comprehensive, overhaul, rewrite, optimize performance, full review, system design, multi-step, parallel, large-scale

Custom keywords

Add project-specific keywords:

[agent.routing]
enabled = true
low_keywords = ["bump version", "update dep"]
high_keywords = ["database migration", "api redesign"]

Length heuristics

• > 200 words — +2 high score (longer prompts tend to describe complex tasks)
• > 80 words — +1 high score

Scoring

The classifier counts keyword matches for low and high categories:

• If high > low → High tier
• If low > high → Low tier
• Otherwise → Medium tier

Per-provider model selection

Each provider maps tiers to specific models via model_for_tier():

OpenAI

Anthropic

Gemini

Events

When routing selects a model, a RoutedModel event is emitted:

RoutedModel { model_name: "claude-haiku-4-5-20250301", tier: Low }

This is displayed in the TUI so you can see which model was selected and why.

Cost implications

Routing can significantly reduce costs by using cheaper models for simple tasks:

• A typo fix routes to Haiku ($0.001) instead of Opus ($0.05)
• A complex refactor still gets the full power of Opus

Track actual costs with nyz cost daily to see the impact.

When to enable

Routing works best when your workflow mixes simple and complex tasks. If you primarily do complex work, the overhead of classification isn’t worth it — just set a high-tier default model in configuration.

Enable routing if:

• You frequently make small fixes alongside large features
• You want to minimize API costs without manually switching models
• You use a provider with clear tier differentiation (Anthropic is ideal)