Research Brief: Anthropic Claude Model Selection & The Advisor Strategy (2026)

• Source Article: *Claude models explained: choosing the best model for your use case* (Anthropic Blog, July 24, 2026)
• Category: Enterprise AI / Model Selection & Multi-Agent Architecture
• File Path: ~/topics/research/Anthropic_Claude_Model_Selection_and_Advisor_Strategy_2026.md

🎯 Executive Summary

Anthropic published its official model selection and workload optimization guide for the Claude 5 generation. Beyond model class breakdowns, Anthropic highlights two major operational principles:

1. "Start Smart" (Cost-Per-Task Economy): Starting with a top-tier model at a lower "effort level" is often cheaper per completed task than starting with a smaller model, because intelligent models require far fewer turns, retries, and context loops.

2. The Advisor Strategy (Sub-Agent Coaching Pattern): Coupling a fast, lower-cost worker model (e.g. Sonnet 5) with a top-tier advisor model (e.g. Fable 5) invoked only for plan verification achieves near-frontier performance at a 37% cost reduction.

📊 Claude Model Family Taxonomy (2026)

┌──────────────────────────────────────────────────────────┐
│ Mythos / Fable (Frontier: Cyber, Bio, Unsolved Workflows) │
├──────────────────────────────────────────────────────────┤
│ Opus (Enterprise Reasoning, Terminal-Bench 2.1)          │
├──────────────────────────────────────────────────────────┤
│ Sonnet (Versatile Workhorse, Sub-Agents, Orchestration)   │
├──────────────────────────────────────────────────────────┤
│ Haiku (High-Frequency, Low-Latency, Utility Routing)    │
└──────────────────────────────────────────────────────────┘

1. Mythos / Fable (Frontier Class):

Mythos:* Restricted to trusted organizations under Project Glasswing for dual-use cybersecurity and biology research.

Fable:* Publicly deployed frontier class with enhanced alignment guardrails.

2. Opus: Heavyweight reasoning and complex agentic coding. Fable provides superior creativity, nuance, and edge-case wisdom when Opus hits quality ceilings on real-world evals.

3. Sonnet: The core operational workhorse for multi-agent subagent delegation and everyday developer tasks.

4. Haiku: High-frequency, low-latency execution for routing, classification, and simple tool calls.

💡 Key Architectural Patterns for Sovereign Agents

1. The "Advisor Strategy" Pattern

• Concept: Run worker subagents on cost-effective, high-speed models (Sonnet 5). Instruct the worker to pause and query a frontier "Advisor" model (Fable 5 / Opus) only at key checkpoints: (a) initial plan validation, and (b) final output verification.
• Benchmark Evidence: On SWE-bench Pro, Sonnet 5 paired with a Fable 5 Advisor scores within 10% of standalone Fable 5, while reducing total token spend to 63% of the full-frontier baseline.

[ User Task ] ──► [ Worker Subagent (Sonnet 5) ]
                          │
                          ├─ (Initial Plan) ──► [ Advisor (Fable 5) ] ──► (Approved Plan)
                          │
                          ├─ (Executes Steps)
                          │
                          └─ (Final Output)  ──► [ Advisor (Fable 5) ] ──► (Verified Result)

2. Cost-Per-Task vs. Price-Per-Token

• High price-per-token models are frequently cheaper on a per-task basis when tasks are multi-turn.
• Smaller models suffer from "retry inflation"—spending tokens on hallucinated function calls, incorrect assumptions, and recovery loops that compound overall cost.

3. Custom Production Evals Over Public Benchmarks

• Standard benchmarks (MMLU, SWE-bench) suffer from saturation on frontier models (Opus / Fable).
• Selection decisions should be driven by custom, production-derived edge-case evals representing real organizational failures.