← All themes and ideas
Credits & caps confidence: medium score 5.3/10 2026-11-10

No measured evidence that Opus-supervising-Sonnet beats Opus-doing-it-all

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

Group member · 2026-11-10  ·  open on Facebook ↗
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.