Chose cheap cloud coding subs over building a local rig

After testing Qwen 3.8 27B and Flash through Qwen's own API, since their own hardware can't run them, this person judged the models capable as sub-agents handed a narrow, well-defined brief, but not for the harder thinking and planning tasks. Weighing a dedicated local rig against staying on cloud, they reason that hardware powerful enough to run those models would cost about as much as a decade of DeepSeek Flash and Mimo 2.5 inference, so they'd rather drop their OpenCode Go subscription and keep just Codex for Sol and Terra at $10 a month. Until hardware prices come down, lighter chatbot work stays on a local Qwen 3.6 35B.