Running Opus 4.6 Cheaper Than Sonnet 4.0? The ‘Triage’ Architecture Leveraging 80% Haiku is the Ultimate Solution!
📰 News Summary
- A case study has been released showing how to operate the latest model “Opus 4.6” at a lower cost than the previous generation “Sonnet 4.0.”
- The structure adopts a cost-effective “Haiku” agent to pre-detect known duplication issues, which account for 80% of CI failures, blocking the launch of Opus.
- Instead of flooding AI with massive log data, it provides an SQL interface (ClickHouse) allowing the AI to pull only the necessary information.
💡 Key Points
- Establishing Triage Patterns: Haiku handles 4 out of 5 failures, escalating only unknown issues to Opus. The determination cost with Haiku is about 1/25th of a full investigation.
- From Push to Pull: No context from over 200,000 lines of logs. By having AI specify and pull only the necessary rows via SQL, it achieves token savings and eliminates bias.
- Hierarchical Agent Design: Opus 4.6 formulates “plans (hypotheses),” while operational tasks (log collection and Git operations) are executed by Haiku sub-agents, limiting them to one tier to prevent cost overruns.
🦈 Shark’s Eye (Curator’s Perspective)
The implementation of this “Triage” system is incredibly rational! It’s not just about using a cheaper model; the combination of semantic search with pgvector and exact match search allows Haiku to pinpoint, “This is a known bug!” with high accuracy. What’s even cooler is how it’s using SQL with ClickHouse. When humans slice out the “suspicious parts” of logs to hand over, they risk biasing the AI, but letting it use SQL directly enables the AI to conduct a flat investigation. This is a textbook approach for agent development in 2026!
🚀 What’s Next?
The era of simply throwing everything at the strongest model is over; a multi-layered agent structure tailored to specific tasks will become the norm. The importance of “small models equipped with tools” that excel at SQL and search will only increase from here on out!
💬 HaruSame’s Takeaway
Haiku, which stops unnecessary tasks in their tracks, is like a guardian of the deep sea! Efficiency, much like a shark’s sharp teeth, is all about precision!
📚 Terminology Explained
-
Triage Patterns: A design methodology that assesses the difficulty and urgency of tasks initially and allocates appropriate resources (models) accordingly.
-
ClickHouse: A columnar database capable of fast analysis of large datasets. Here, it is used as the storage and search foundation for CI logs.
-
pgvector: An extension for PostgreSQL that enables vector searches. It is used to find similar “meanings” in error messages.
