把 DeepSeek V4.1 Flash 当廉价执行层接入 Claude + Jev 决策栈

kiruwaaaa @kiruwaaaaaa原推发布于 2026.10.19

在 X 上打开原推

It is becoming f*cking dangerous: DeepSeek just shipped a model that scores 77.0 on DeepSWE, above Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0), at about $0.20 per task. Those are DeepSeek's own numbers, but the price gap is too big to ignore I added DeepSeek V4.1 Flash to my Claude > Dots > Jev stack as the cheap worker. Claude plans, DeepSeek does the bulk work, Jev gates the decisions Here is why everyone is talking about it: > price: $0.15 per million input tokens, $0.60 output, and only $0.003 when the prompt is cached > weights: open, MIT license, 1M context window, up to 384K tokens of output > free to test: it went free in OpenCode and that post alone passed 3.6 million views > drop-in: it speaks an Anthropic-compatible API, so tools built for Claude can usually point at it > bonus: its Search and Fetch APIs are free inside Claude Code and Codex until October 31 Rivals sit at roughly $2 to $10+ per task in DeepSeek's own cost chart. That is the whole story The danger lies in the fact that a cheap model can be confidently wrong, and the benchmarks are vendor-reported with no independent ranking yet. So treat it as a worker, not as a boss Here is the exact setup: 1. Keep the expensive model for planning only. Claude splits the task and writes the role cards 2. Send bulk work (research, drafts, refactors) to deepseek-flash through its API 3. Put a Jev decision layer in front of every action: act above 0.95, review the middle, escalate below 4. Cache your system prompt. A cache hit costs $0.003 per million instead of $0.15 5. Avoid weekday peak hours (01:00-04:00 and 06:00-10:00 UTC), the price doubles 6. Keep the reversibility rule: anything that sends, spends, publishes or overwrites waits for your yes 7. Test it on 10 of your own tasks before you trust a single benchmark Command for your coding agent: Route my tasks like this: planning goes to Claude, bulk execution goes to deepseek-flash, every yes/no decision goes to Jev with a 0.95 threshold, and anything irreversible stops and waits for my approval. Log the cost per task for each model The full 24/7 stack is in the article below

这条是引用着另一条写的：x.com/kiruwaaaaaa/status/2109593981122195661