Model Router

When to Override

The routing table provides defaults. Override when:

User Preference

The user explicitly requests a specific model for a task: "Use Opus for drafting this section -- the tone is critical." Always respect explicit model requests.

Cost Constraints

When the user has budget constraints or the session involves many sub-agent dispatches:

• Downgrade Opus tasks to Sonnet for Tier 1 (exploratory) verification -- spot-checking 3-5 values doesn't need the strongest model.
• Batch multiple Sonnet tasks into fewer dispatches to reduce overhead.
• For very large documents with 50+ claims, consider running initial verification with Sonnet and escalating only flagged claims to Opus.

Speed Requirements

When the user needs results quickly:

• Use Sonnet for all tasks except the Integrity Gate (which should always use Opus or be performed by the orchestrator directly).
• Skip parallel review dispatches and have the orchestrator do inline review.

Complexity Escalation

When a Sonnet agent returns uncertain or conflicting results:

• Re-dispatch the same task to Opus with the Sonnet output as additional context.
• This "escalate on uncertainty" pattern catches the cases where the lighter model's limitations matter.

Small Context Tasks

When the task is small enough that the orchestrator can handle it inline (a single paragraph edit, one value lookup), skip the sub-agent dispatch entirely. The overhead of creating a sub-agent prompt, dispatching, and parsing results isn't justified for trivial tasks.

Model Capabilities Relevant to Document Work

Opus

• Strongest analytical reasoning -- catches subtle logical errors, inconsistent claims, and unsupported conclusions
• Best at cross-referencing multiple files and holding complex relationships in working memory
• Most reliable for mathematical verification (though still verify independently)
• Better at understanding domain conventions and compliance requirements
• Higher cost, slower response time
• Context window: the orchestrator typically runs on Opus with a large context (up to 1M tokens)

Sonnet

• Fast, fluent prose generation that matches existing document tone
• Strong pattern matching for structural analysis and extraction
• Good at detecting AI-telltale patterns in text (parallel structures, excessive hedging)
• Efficient for broad research across large documents
• Lower cost, faster response time
• Context window: 200K tokens, sufficient for focused single-task packets

Dispatch Annotation

When dispatching a sub-agent, annotate the intended model in the prompt contract:

[Model: Opus]
Task: Verify all quantitative claims in Aim 1 against source data.
...

[Model: Sonnet]
Task: Draft the revised Background section using verified values from the Claim Ledger.
...

If the runtime environment doesn't support model routing (e.g., all sub-agents run on the same model as the orchestrator), the annotation is informational -- it documents the intended cognitive tier for the task, which helps the orchestrator calibrate expectations for the output quality.

Cost-Effectiveness Guidelines

For a typical document editing session with the agentic loop:

This gives a typical ratio of approximately 60% Sonnet dispatches, 40% Opus dispatches by count, with the Opus dispatches concentrated at the verification and review checkpoints where accuracy matters most.