9 hours ago
- Running a coding agent locally requires expensive VRAM (e.g., NVIDIA 5090 or 2×3090s) for decent performance.
- Single-stream development yields ~2k tok/s input and 150 tok/s output, costing about $0.80 per hour on OpenRouter for Qwen 3.6.
- Agent swarms (e.g., 32 concurrent agents) drastically increase token consumption but also make local LLM rigs cost-effective compared to API rates.
- Cost comparison: local rig electricity cost ~$0.11 per hour for swarm vs $7.06 on OpenRouter, and $89 on Sonnet 5.
- Over a month of 8h/day, local rig costs ~$26 vs $1,700 on OpenRouter and $21,400 on Sonnet 5 for swarm.
- Agent swarms shift tokenomics to favor local AI by saturating GPUs and reducing reliance on expensive API calls.