Advisor Strategy: Opus as Advisor Saves 12%+ on Agents

Pair cheaper Haiku or Sonnet as executors with Opus as advisor for near-Opus performance: Sonnet+Opus boosts SWE-bench by 2.7 points and cuts agentic task costs 12%; Haiku+Opus doubles browse-comp score from 19.7% to 41.2% while staying cheaper than solo Opus.

Advisor Strategy Delivers Opus Intelligence at Fraction of Cost

Anthropic's advisor strategy routes tasks to cheap executors (Haiku at $1/M input, $5/M output; Sonnet at $3/$15) that call Opus ($5/$25) only for hard steps, avoiding waste on simple ones. In evaluations, Sonnet+Opus advisor raises SWE-bench scores 2.7 points over solo Sonnet while reducing agentic task costs nearly 12%. Haiku+Opus jumps browse-comp from 19.7% (solo Haiku) to 41.2%, still cheaper than solo Opus. Output tokens cost 3-5x more than input across models, so advisor minimizes expensive Opus outputs. Limit advisor calls with max_uses in API requests for cost control.

Dashboard Tests Show Quality Holds on Easy-to-Hard Prompts

Build a frontend dashboard calling Messages API to compare setups: Haiku+Opus, Sonnet+Opus, Sonnet solo, Opus solo. Easy prompts (e.g., "business hours?") stay with executor—no advisor call—costing 21x less than Opus solo (Haiku: ~$0.006 vs Opus: ~$0.13), with similar accuracy. Medium prompts like integrations use knowledge-base search without advisor. Hard ones (e.g., "return hardware+software after 3 weeks?") trigger Opus: Haiku+Opus nails 30-day window, original packaging, non-refundable software, creates ticket—better than Sonnet solo's vague sales handoff. Sonnet+Opus gives polished customer response but costs more than Haiku+Opus. Opus solo matches advisor quality on complex tasks but inflates costs unnecessarily. Test 100s of prompts per use case; advisor excels when executor self-assesses need accurately.

Apply in Claude Code with Opus Plan Mode to Extend Sessions

Claude Code lacks native advisor (API-only via /v1/messages endpoint with type: advisor-20260301), but mimic via /model opus plan: plans in Opus 4.6, executes in Sonnet 4.6 (or Haiku for summaries/searches). Status bar shows mode switch. Demo: Opus-plan+Sonnet builds clearer advisor dashboard (dynamic elements, savings sliders) vs Opus-only's static version—same quality, lower session burn (Opus eats limits 2-3x faster). Use Sonnet default, Opus plan for planning/alignment, Haiku for exploration. Messages API suits custom apps/automations (tools, images, stateless); Claude Code for local file edits/terminal.

Trade-offs: Test Extensively Before Production

Advisor shines for agentic workflows with mixed difficulties—smart escalation beats brute-force Opus. But executors may skip advisor on borderline tasks (Haiku misses some Sonnet flags), so benchmark your prompts: advisor+ consistently tops solo mid-tier models, matches Opus on hard. Not for every task; prototype via shared GitHub repo (free Skool community). Beta feature—monitor for refinements like new monitor tool.

Video description
Full courses + unlimited support: https://www.skool.com/ai-automation-society-plus/about?el=advisor-strategy All my FREE resources: https://www.skool.com/ai-automation-society/about?el=advisor-strategy Apply for my YT podcast: https://podcast.nateherk.com/apply Work with me: https://uppitai.com/ My Tools💻 14 day FREE n8n trial: https://n8n.partnerlinks.io/22crlu8afq5r Code NATEHERK to Self-Host Claude Code for 10% off (annual plan): https://www.hostinger.com/vps/claude-code-hosting Voice to text: https://ref.wisprflow.ai/nateherk Anthropic just released the advisor strategy, and it's a big deal for anyone building with Claude. Instead of running every task through Opus, you can pair it as an advisor with a cheaper model like Sonnet or Haiku doing the actual work. In this video I built a dashboard to test how it performs across easy, medium, and hard prompts, and I also show you how to use the same idea inside Claude Code with opus plan mode to stretch your session limit further. Sponsorship Inquiries: 📧 sponsorships@nateherk.com TIMESTAMPS 0:00 What Is the Advisor Strategy 1:34 Model Pricing Breakdown 2:14 Messages API vs Claude Code 4:11 Dashboard Demo Starts 9:34 When to Use It 10:24 Advisor Strategy in Claude Code 11:45 Opus Plan vs Opus Only Build 14:28 Final Thoughts

Summarized by x-ai/grok-4.1-fast via openrouter

7857 input / 1371 output tokens in 13702ms

© 2026 Edge