Model Routing

IO uses dynamic model routing to balance cost and capability. There are two contexts: the orchestrator and squad agents.

Orchestrator

The orchestrator always uses the model specified in your config (defaultModel). It does not dynamically switch models per message.

{
  "defaultModel": "gpt-4o"
}

If not set, it defaults to gpt-4o.

Squad Model Selection

When a Team Lead assigns tasks to agents, it uses an LLM call to classify task complexity and select the cheapest capable model.

How It Works

1. The task description is sent to the cheapest available model (trivial tier)
2. That model classifies the task into one of 5 tiers
3. The cheapest available model in the classified tier is selected
4. If the selected model fails, it automatically retries with the next tier up (max 1 escalation)

Tiers

Models are automatically assigned to tiers based on their Copilot premium request multiplier:

Tier	Multiplier Range	Example Models
Trivial	≤0.33x	gpt-4o-mini, gpt-4o, gpt-5-mini
Fast	0.34x–1x	Gemini 2.5 Pro
Standard	1.1x–5x	GPT-5.1, GPT-5.2
Premium	5.1x–15x	GPT-5.4, Claude Sonnet 4.5/4.6
Ultra	15x+	Claude Opus 4.6/4.7/4.8, GPT-5.5

Dynamic Model Catalog

IO fetches available models from GitHub's Models API and scrapes pricing from GitHub's documentation. This happens:

• At startup (blocks until complete — ensures models are always available)
• Periodically on the configured pricingRefreshHours interval (default: 24 hours)

If all sources fail, IO falls back to hardcoded seed data for known models.

Cost Tracking

Every LLM call records two cost metrics (point-in-time pricing):

• Premium Request Cost — the model's Copilot premium request multiplier at time of use
• Token Unit Cost — computed as (input_tokens × input_multiplier + output_tokens × output_multiplier) × $0.00001

Both are stored per usage record and never recalculated when pricing changes. Historical costs reflect the rates when the call was made.

Configuration

{
  "defaultModel": "gpt-4o",
  "pricingRefreshHours": 24
}

Field	Default	Description
defaultModel	gpt-4o	Model used by the orchestrator for all messages
pricingRefreshHours	24	How often to refresh the model catalog and pricing data