One R9700 running Qwen3.6-35B did $426.80 of equivalent Opus coding work in a week
supracode · reddit · 2026-10-15
A Redditor details a week of agentic coding on a single AMD R9700 running Qwen3.6-35B-A3B (UD-Q6K, 220k context) with the Claude CLI: 46 sessions, 20.7M input tokens, 3.2M output, and 1.4B cache reads — about $426.80 in equivalent Opus 5.5 API pricing, with cache reads making up 65.6% ($9.28/session). Caveat: Qwen gets confused with workflow commands at long context, so the author pays for ChatGPT as a second reviewer and may subscribe to Opus for code review and planning. The takeaway: heavy cache-reuse workloads make local GPUs economically attractive, though frontier models still win on quality.
More from coding & agent
- Benchmarking coding agent harnesses and inference: lessons from $100Ks/month in research spend — abhijithneil · 2026-10-15
- AssetFurnace: Free Open-Source Local Image-to-3D Pipeline for Game Characters — chongdashu · 2026-10-15
- BaoCut's open-source approach to AI bilingual subtitle alignment, in five pitfalls — dotey · 2026-10-15
- Claude Code ships 278 changes in 2 days: 7 updates worth knowing — minchoi · 2026-10-15
- Video edited entirely by Claude Opus 4.5 High using free self-made skills, no editing software — chongdashu · 2026-10-15
- How to Connect Grokbot to Claude Code, Codex or Any AI Agent via HQ Messages — jacob_posel · 2026-10-15