Model Selection

Overview

RelayPlane analyzes each incoming request and classifies it by complexity. Simple tasks get routed to cheaper, faster models while complex tasks use premium models. This happens transparently — your tool doesn't need to change anything.

Task Classification

The proxy infers task type from the request content using pattern matching on the last user message. It recognizes 9 task types:

Complexity Scoring

Each request is scored for complexity based on the last user message (not system prompts or conversation history). The scoring considers:

• Code indicators — code blocks, function/class definitions (+2)
• Analytical tasks — analyze, compare, evaluate, review (+2)
• Math/logic — calculate, solve, prove, derive (+2)
• Multi-step reasoning — step-by-step instructions (+2)
• Architecture/design — system design, distributed systems (+3)
• Implementation requests — implement, refactor, debug, optimize (+2)
• Token length — longer messages score higher (+1 to +4)
• Multiple requirements — many "and" conjunctions (+1 to +2)

The score maps to three complexity levels:

• Simple (score < 2) → cheapest model
• Moderate (score 2-3) → mid-tier model
• Complex (score ≥ 4) → premium model

Default Model Mapping

The default complexity-to-model mapping (configurable in ~/.relayplane/config.json):

1{
2  "routing": {
3    "complexity": {
4      "enabled": true,
5      "simple": "claude-sonnet-4-6",
6      "moderate": "claude-sonnet-4-6",
7      "complex": "claude-opus-4-6"
8    }
9  }
10}

Routing Modes

You can request a specific routing strategy using model aliases:

Cascade Mode

When routing.mode is set to "cascade", the proxy starts with a cheaper model and escalates to a more expensive one if the response shows signs of uncertainty or refusal:

1{
2  "routing": {
3    "mode": "cascade",
4    "cascade": {
5      "enabled": true,
6      "models": ["claude-sonnet-4-6", "claude-opus-4-6"],
7      "escalateOn": "uncertainty",
8      "maxEscalations": 1
9    }
10  }
11}