Budget agent stack: Claude Code on DeepSeek with Jev for small calls
A five-step setup routes Claude Code to deepseek-flash and sends small decisions to Jev at $0.042 per 1M tokens, about $0.20 per task.
a Chinese schoolkid with $5 can now outbuild your whole f*cking AI agent stack. here's the exact setup step 1 → point Claude Code at DeepSeek's Anthropic endpoint and set EVERY model name to deepseek-flash. skip one and "opus" quietly goes to V4 Pro step 2 → send every small call to Jev. "which file?" "retry or stop?" "ship it?" $0.042 per 1M tokens, output free step 3 → one long session per project. cached input is $0.003 per 1M, so rereading your repo is basically free step 4 → 2 tries per fix, then park it. heavy runs wait for off-peak, half the price step 5 → the expensive model reads only the final diff. nothing else Flash-1017 is about $0.20 per DeepSWE task on DeepSeek's own chart. Opus 5 is around $10. that's 50 Flash tasks for one no team. no funding. just rules one catch: the chart compares Opus 5, not 5.5. and Flash is cheaper, not faster so who's still paying the $10 version for every step?
This post quotes another post: x.com/tekari0/status/2111765959342195160