Dev Tools · 1h ago
Free Local AI Executor Costs More Than Cloud-Only Opus in Coding Tests
A developer ran 40 trials comparing agentic coding configurations and found that using a free local Qwen model as executor under Opus orchestration was the most expensive cloud option. The orchestrator's prompt-cache re-reads of Qwen's summaries caused Opus to consume 1.4–5.3x more tokens than Opus solo. Haiku solo was 5.5x cheaper than Opus solo on the largest task but failed 25% of the time.
Meridian48 take
The assumption that local execution is always cheaper ignores the hidden cost of orchestrator token consumption, a lesson for anyone building multi-model agent pipelines.
Read the full reporting
When the Free Executor Cost More: 40 Trials on Opus + Local Qwen Ended Up the Most Expensive Cloud Arm →
DEV Community
agentic-codingllm-cost-analysis