实习生用本地开源模型替掉每月 3.8 万美元的 GPT-6 Astra 账单
帖主称实习生一个周末把内部工具全指向本地模型，账单从 3.8 万美元降到 260 美元，并质疑 Astra 的定价与实际能力。
openai billed our startup $38,000 last month. our intern erased the whole thing over a weekend with a free model, and the senior engineers who spent a year saying it could not be done have gone very quiet. he had no permission and no budget, he just pointed everything at a local model and let the results talk. gpt-6 astra was doing nothing his laptop could not, we were paying $38k a month for the logo on the invoice. here is exactly what a 21 year old did in a weekend, and it runs for about $6 instead of $400: -> he pulled an open model like grok or kimi k3 and served it on one nvidia card the office already owned -> he pointed cursor, cline and every internal tool at localhost, and 4,000 daily jobs never noticed the switch -> the heavy stuff went to a gpu rented by the hour, the rest ran local for the cost of power -> the $38k invoice came back at $260 the next cycle, and nobody in the room could argue with the number here is the part they will fight me on: openai and anthropic count on switching sounding like a quarter-long project. an intern proved it is an afternoon. the $38k was never buying capability, it was buying the belief that this was hard. the full price map is in the article below, gpt-6 astra and fable 5.1 sitting at $10/$50 while kimi k3 stays at $3/$15. read it before your own invoice does the math for you like it did ours.
这条是引用着另一条写的：x.com/starmexxx/status/2098293921994465300