model-router

Model Router

Route each sub-agent task to the Claude model best suited for its cognitive demands. The goal is cost-effective model selection: use the strongest model where accuracy is non-negotiable (verification, review, complex reasoning), and use the faster model where fluency and throughput matter more than deep analysis (research, drafting, extraction).

Routing Table

Task Type	Model	Rationale
Source verification	Opus	Cross-referencing values across files, resolving discrepancies, and catching subtle errors requires deep analytical reasoning. A missed value here propagates through the entire document.
Derived calculation verification	Opus	Independent recomputation of fold-changes, ratios, and percentages. Math errors compound, so this needs the most capable model.
Domain review (accuracy, compliance)	Opus	Assessing whether claims are supported by evidence, whether domain conventions are followed, and whether compliance requirements are met. Requires domain knowledge and careful judgment.
Citation integrity check	Opus	Verifying that each citation supports the claim it's attached to after restructuring. Requires understanding the semantic relationship between claim and source.
Integrity Gate checks	Opus	Final verification pass. The gate is the last line of defense -- it must catch what earlier steps missed.
Research and exploration	Sonnet	Broad document reading, inventory building, extraction of structure and content. Speed and coverage matter more than deep analysis.
Drafting	Sonnet	Prose generation with verified values and citation maps. The values are already verified; the drafting agent's job is fluent writing, not fact-checking.
Language review (tone, naturalness)	Sonnet	AI-telltale detection, tone consistency, style evaluation. Pattern recognition and stylistic judgment, not deep reasoning.
Structure planning	Sonnet	Outlining section movements, content consolidation maps. Organizational work that benefits from speed.
Stale value scanning	Sonnet	Searching documents for deprecated values. Pattern matching, not analytical reasoning.
Format and style checks	Sonnet	Heading hierarchy, word counts, formatting compliance. Mechanical checks that don't require deep understanding.

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:

Phase	Typical Dispatches	Model Mix
Research	1-2 explore agents	Sonnet
Source Verify	1 verification agent	Opus
Structure Plan	0-1 (usually inline)	Sonnet or inline
Draft	1-3 drafting agents	Sonnet
Apply Gate	0 (inline)	Orchestrator
Integrity Gate	1 verification agent	Opus
Dual Review	2 review agents	Opus (domain) + Sonnet (language)
Fix	0 (inline)	Orchestrator

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.