Running Opus 4.6 costs less than Sonnet 4.0 because a Haiku triager filters 80% of CI failures before they reach the frontier model.
Key Takeaways
Haiku triager with pgvector semantic search screens duplicates; 4 of 5 CI failures never reach Opus, at 25x lower cost per triage.
Opus never reads raw logs; it queries ClickHouse via SQL and spawns scoped Haiku sub-agents capped one level deep to prevent runaway costs.
Haiku handles 65% of input tokens but only 36% of LLM spend; removing the model hierarchy more than doubles the daily bill.
Context hygiene keeps Opus clean: sub-agents return structured summaries, not raw output; each sub-agent starts fresh and its context is discarded after.
The triager/orchestrator pattern generalizes to security logs, IoT telemetry, and financial data wherever most events are noise or known repeats.
Hacker News Comment Review
The sole commenter flagged the title as misleading clickbait; the actual insight is the Haiku-triager-plus-Opus-orchestrator architecture, not the counterintuitive cost headline.
