Claude 额度差距来自配置：用 DeepSeek V4.1 Flash 当廉价子代理

Om Patel @om_patel5原推发布于 2026.10.26

在 X 上打开原推

same claude plan. one person burns the 5-hour limit in 15 minutes. another runs 5 sessions almost 24/7 and lets the reset expire unused. it's not the model. i went through this week's biggest token and claude.md threads, plus the guides and papers people kept linking, and the gap is almost all setup. - cheap hands. one person found all 12 subagents were running opus. after telling it "opus orchestrates, subagents use cheaper models", the same prompt burned 10% of the usage. - open source workers. someone pushed 66 million tokens through deepseek v4.1 flash and the bill was $0.55. another runs codex and deepseek as subagents for about €50 a month. - fewer tools. anthropic's own guide says tool choice gets worse past 30 to 50 tool definitions. they cut claude code's system prompt by over 80%. - no blind searching. one line, "no system-wide greps or finds, ask if you don't know where something is", nearly doubled how many sessions one dev can run at once. - read once, edit once. "list every file, read them in one pass, make all edits, run tests once" killed a read-edit loop that was eating half one person's usage. - code over docs. a paper out this week (320 runs, claude code and codex) found agents follow a stale claude.md over the actual code 90% of the time. one sentence saying the code wins moved that from 4.2% to 91.4%. here's what i think those threads are missing. open source isn't replacing claude. it's becoming the layer underneath it. the expensive model decides, the cheap ones type. same split every company already has between a senior engineer and the team. so the skill that matters now isn't prompting. it's budgeting. which model gets which job, what loads into context, what never loads at all. the people out of tokens by 11am aren't using too much ai. they're paying senior rates for junior work. and almost every fix on that list is a deletion. fewer tools, fewer files read twice, a shorter claude.md, no grepping the whole disk. the best setups this week got better by taking things out. my guess: within a year, "how do you route your models" is a normal interview question, the same way "how would you cut this aws bill" is now.