下面这篇来自我一个朋友。他在一家公司做 CTO，上个月跟我喝酒，倒了一晚上苦水。我劝他写下来，他懒，我就替他整理了这一版，口气照他的。

给团队开了 100 万人民币的 token 不限量，我学到的最贵的一课

我跟老板申请了 100 万人民币，给我那 20 多人的团队搞 token 不限量。现在这笔钱烧得差不多了，回头看，这大概是我今年最贵的一次交学费。

当初提这方案的时候，我把 PPT 做得可漂亮了：”打破资源束缚，释放团队潜能”。我们 CFO 盯着我问了一句”你确定不是给模型厂商送钱？”，我当时还笑他保守。

呵呵。

先说人。我原以为敞开用 AI，大家能早点下班。结果呢？上个月我半夜两点收到一个工程师的消息，说”哥，我睡不着，五个窗口还开着，AI 在跑，我不敢关”。中位数的人同时开 5 个 session 等推理，脑子在不同 context 之间反复横跳。人脑不是为这个设计的——你让它同时跟五个对象聊天还得记住每个人的前情提要，迟早崩。更损的是，AI 哗哗吐东西，最后还得人去 review。我之前以为把写活甩给 AI 是解放，其实是把”写”换成了”审”，活没少，还多了层”怕它写错”的焦虑。

再说钱。单日 token 花得最猛的成员，一天烧掉 5 万。团队中位数是 1.4 万一天。大家最爱用的俩模型是 GPT-5.6 Sol Fast 和 Fable 5 Fast。把这些数字跟工资一比就明白了——很多人一天的 AI 账单，比他一个月工资还高。这账没法持续。很多人图省，把快问快答、前期调研丢给 Grok 4.5，便宜嘛，快嘛。等是短了，可来回拉锯上来了。贵模型一次能搞定的事，它迭代五轮。再加上那五六个并行的 session，一折算，有时候比直接用贵模型还贵。

便宜，是个幻觉。

我后来才想明白，模型路由这事儿没那么”贵的留给难的”——难的它未必接得住，简单的它又磨叽。这 100 万里，至少有一半是交了”以为省了其实没省”的学费。

最让我后背发凉的是第三个。上个月我们团队会议数环比掉了快 80%。表面看是好事，少开会嘛。可我扒开一看，不是人少开会了，是变成 AI 跟 AI 在聊。好些人把脑子托管出去了，项目细节说不清。上周三一个跨组 sync，我问我这人进度，他沉默三秒：”我让 AI 问 AI 了，稍等。”那一刻我有点慌——答案在 AI 那头，人成了传声筒。还有更离谱的：群里 A 的 AI 乱答一通，对面 B 的 AI 再乱答一通，俩人隔空把天聊飞了，主题早没影了，全程没一个真人在掌舵。

所以结论很清楚，但不好听。

无限 token 不切实际，它比雇个真人大概要贵出一个量级，还在组织效率上撞到了边际效益的天花板。

不是 token 不够。是公司这套架构从根上还是为人搭的。我们嘴上喊 AI-native，身体很诚实——汇报线、会议、KPI、考核，全是围着人转。你给 AI 放开用，它水土不服，因为这些流程压根没给”一个人指挥一群 AI 并行”留座位。

我甚至开始怀疑，现在满世界喊”AI 降本增效”的那帮人，是不是都没真带过队。降本？我这边成本翻倍了。增效？人更累了。真增效的是模型厂商的财报。

至于这 100 万到底买到了什么，我到现在也没算清账。买到一个教训，大概值这个价。那个实验我不续了——但团队的 token 我也不会掐死，只是从”敞开喝”改成”看着喝”。

PS：写这篇的时候，我发现自己不知不觉又开了四个窗口。老毛病。

English Version

This one comes from a friend of mine. He’s a CTO at a company, and last month over drinks he dumped a whole evening of complaints on me. I told him to write it down; he was too lazy, so I cleaned it up in his voice.

I talked our boss into ¥1 million for my 20-odd-person team to use AI tokens with no limit. The money’s mostly burned now, and looking back, that was probably the most expensive tuition I paid this year.

When I pitched it, I made a beautiful slide deck: “Break resource constraints. Unleash team potential.” Our CFO stared at me and asked, “You sure you’re not just laundering money to the model vendors?” I laughed at him for being conservative.

Ha.

Let’s talk about the people first. I figured if everyone could use AI freely, they’d go home earlier. Right. Last month an engineer messaged me at 2 a.m.: “Bro, can’t sleep, five windows still open, AI’s running, I’m afraid to close them.” The median person keeps five sessions open at once, waiting on inference, their brain jumping between contexts. Human brains were never built for that—it’s like dating five people simultaneously and remembering each one’s backstory. Something breaks. And then the AI floods out output and a human still has to review it. I thought offloading the writing to AI was freedom. Really I just swapped “writing” for “proofreading,” and added a fresh layer of anxiety about whether the thing is wrong.

Now the money. The heaviest token spender burned ¥50,000 in a single day. The median was ¥14,000 a day. The two favorite models were GPT-5.6 Sol Fast and Fable 5 Fast. Line those up against salaries and it’s obvious—for a lot of people, one day of AI bills costs more than a month of their paycheck. That math doesn’t survive contact with reality. A lot of people threw quick Q&A and early research at Grok 4.5 because it’s cheap and fast. The wait’s shorter, sure—but the back-and-forth multiplies. Things the expensive model finishes in one shot, the cheap one iterates through five. Add the five parallel sessions on top, convert it all to price, and sometimes it costs more than just using the expensive model.

Cheap is a hallucination.

It took me a while to see that model routing isn’t “save the pricey one for the hard stuff.” The hard stuff it can’t always handle; the easy stuff it drags its feet on. At least half of that million was tuition for “thought I saved, didn’t.”

The one that chilled me was the third. Last month our team’s meeting count dropped nearly 80% month over month. Looked great on paper—fewer meetings! But when I looked closer, it wasn’t that people met less. It was AI talking to AI. A lot of folks had effectively outsourced their brains, and they couldn’t explain project detail anymore. In a cross-team sync last Wednesday, I asked one of my guys for progress. He paused three seconds: “I had my AI ask the other AI, hold on.” That’s when I got nervous—the answer lived in the AI, and the human had become a Relay. Worse: on a group chat, A’s AI rambled an answer, B’s AI rambled back, and the two of them drifted the whole conversation off the map. No human at the wheel.

So the conclusion is clear, and it’s not flattering.

Unlimited tokens aren’t practical. They cost an order of magnitude more than hiring a real person, and they hit a hard ceiling on organizational efficiency.

It’s not that we ran out of tokens. The company’s org structure was designed for humans, root and branch. We pay lip service to AI-native, but the reporting lines, the meetings, the KPIs—all of it orbits the human. You let AI run free and it fits badly, because none of those processes made room for “one human directing a swarm of AIs in parallel.”

I’ve started to suspect the people shouting “AI cuts costs and boosts efficiency” have never actually run a team. Cut costs? My costs doubled. Boost efficiency? People are more tired. The only thing that got more efficient was the model vendors’ earnings report.

What that million actually bought, I still haven’t tallied. A lesson, maybe, and that’s probably worth the price. I’m not renewing the experiment—but I’m not choking off the team’s tokens either. Just switching from “drink all you want” to “drink where I can see you.”

PS: while writing this, I noticed I’d somehow opened four windows again. Old habits.