The need
A user runs a workflow where Opus directs Sonnet to write code and then review, but the community can only offer opinions — nobody has a measurable use case proving the cheaper pipeline matches full-Opus quality.
Original post — verbatim Thai
สอบถามครับ ระหว่างคนที่ทำ workflow ให้ตัว Opus สั่ง Sonnet เขียนโค้ด แล้วตรวจ กับ Opus ทำเองทั้งหมด มี usecase ที่วัดผลได้มั้ยครับ
Three business opportunities
OPPORTUNITY 01
Model-combo benchmark harness
- What
- An open test suite of realistic coding tasks that scores quality vs spend for pipeline configurations (Opus→Sonnet vs single model) and publishes a live leaderboard.
- Who pays
- Developers and team leads pay for pro runs on their own tasks; the leaderboard itself stays free.
- Price
- ฿290/mo pro harness, free leaderboard (hypothesis)
- Why now
- Model routing is the main cost lever in 2026 Claude subscriptions and every Thai user guesses; harness engineering is now cheap with agents.
- Effort
- M
- Moat
- Longitudinal results across model generations become a reference dataset nobody else has.
- Risk
- Fast model churn invalidates results quarterly, eroding trust in the numbers.
OPPORTUNITY 02
Pipeline cost radar for Claude Code
- What
- A CLI overlay that logs per-task token spend across orchestrated models, flags waste and suggests routing policies ('this file type, use Sonnet').
- Who pays
- The same heavy users, but paying for ongoing spend visibility rather than one-off benchmark runs.
- Price
- ฿199/mo per workspace (hypothesis)
- Why now
- Weekly caps make spend-visibility urgent; existing telemetry is raw JSON nobody reads.
- Effort
- S
- Moat
- Each user's historical traces personalize recommendations — better with every session.
- Risk
- Anthropic ships native cost dashboards.
OPPORTUNITY 03
Workflow tuning audit service
- What
- Done-for-you audit: an expert restructures a client's Claude Code setup (model assignment, review loops, caching) and reports measured savings.
- Who pays
- Freelancers and small teams pay a fixed fee for guaranteed savings without buying or running software.
- Price
- ฿2,500 per audit (hypothesis)
- Why now
- Post-capping economics in 2026 make 'spend less, same output' a sellable promise; AI tooling makes one auditor serve ten clients a month.
- Effort
- S
- Moat
- Before/after case studies compound into pricing power.
- Risk
- Results vary by codebase; one bad audit kills reputation in a small community.