Meta 启动 AI 使用限额管理:终结内部＂Claudeonomics＂竞赛,推广自研 MetaCode

点击蓝字关注我们

报道核心内容中英文对应翻译

On June 13, 2026, Meta sent an internal memo to roughly 6,000 employees. The message was blunt: AI usage caps are here. Budget limits. Token consumption controls. The company's self-built AI Gateway platform will now monitor every team's AI spend — and when token expenditure spikes, the system fires off automatic alerts. The memo concedes that internal AI use is growing so fast that the tools employees lean on this year alone could cost tens of billions of dollars. It's a sharp U-turn: Meta once cheered AI adoption; now it's clamping down, mirroring a broader cost-cutting wave sweeping American enterprises.

2026 年 6 月 13 日，Meta 向大约 6000 名员工发出内部备忘录，措辞直白：AI 使用上限来了，预算限制来了，Token 消耗管控也来了。公司自研的 AI Gateway 平台全面上线，实时监控每个团队的 AI 花销——Token 支出一旦异常飙升，系统自动报警。备忘录承认，内部 AI 用量增长迅猛，仅今年员工使用的工具就可能烧掉数百亿美元。这是一次急转弯：曾经力推 AI 的 Meta，现在开始勒紧缰绳，也折射出美国企业降本增效的大潮。

· · ·

How did we get here? Meta lit the fuse itself. In November 2025, Chief People Officer Janelle Gale declared that "AI-driven impact" would be a core performance metric starting 2026. By January, the company had overhauled its review system, dangling bonuses of up to 200% for top performers. The perverse result: an internal leaderboard called "Claudeonomics" — a cheeky nod to Anthropic's Claude model — tracking tokens burned by more than 85,000 employees. The top 250 got ranked. Winners earned gamified titles like "Cache Wizard," "Token Legend," and "Session Immortal."

这一切是怎么开始的？Meta 自己点的火。2025 年 11 月，首席人事官 Janelle Gale 宣布 "AI 驱动的工作成果" 将成为 2026 年核心考核指标。今年 1 月，公司全面改革绩效体系，最优者奖金高达 200%。扭曲的激励随之而来：一个名为 "Claudeonomics" 的内部排行榜悄然上线——名字借自 Anthropic 的 Claude 模型——追踪超过 8.5 万名员工的 Token 消耗。前 250 名上榜，头衔五花八门："缓存法师"（Cache Wizard）、"Token 传奇"（Token Legend）、"会话不朽者"（Session Immortal）。

· · ·

The numbers got absurd fast. In 30 days, employees collectively torched 60 trillion tokens. The top individual alone: 281 billion — 9.4 billion a day. For context, a typical Claude conversation chews through 1,000 to 10,000 tokens. Some staffers instructed AI agents to run multiple tasks for hours, just to pump up the count. Even Mark Zuckerberg and CTO Andrew Bosworth didn't crack the top 250. At public API rates, that 60-trillion-token binge was worth an estimated $9 billion. When The Information exposed the dashboard in April 2026, Meta killed it within days. The shutdown message read: "It was meant to be a fun way for people to look at tokens, but due to data from this dashboard being shared externally, we've made the decision to shutter Claudeonomics for now."

数字很快变得荒诞。30 天内，员工集体烧掉 60 万亿 Token。排名第一的个人用户一人就消耗了 2810 亿——每天 94 亿。什么概念？一次普通 Claude 对话用掉 1000 到 10000 个 Token。有的员工让 AI 代理同时跑多项任务，一跑就是几小时，纯粹为了刷量。连扎克伯格和 CTO 博斯沃思都没挤进前 250。按公开 API 价格计算，这 60 万亿 Token 价值约 90 亿美元。2026 年 4 月，The Information 曝光了该排行榜，Meta 数日内紧急下线。关停页面留下一句话："本意是让人们以一种有趣的方式查看 Token 使用情况，但由于数据被外部传播，我们决定暂时关闭 Claudeonomics。"

· · ·

Meta CTO Andrew Bosworth had tried to rein things in earlier. In an April memo, he warned: "Nobody should use AI just for the sake of using AI." Token consumption, he stressed, "is not a measure of impact in any sense." Words alone couldn't stop the hemorrhage. Now Meta is building a system. Product developers and engineers have put together a real-time dashboard — AI Gateway — that tracks company-wide AI usage and spending, with automatic alerts on consumption spikes. The company plans to roll these controls out to more employees in the coming weeks and push toward structured token budgets by 2027.

CTO 安德鲁·博斯沃思其实早就试图踩刹车。4 月份的一份备忘录里，他警告："没有人应该为了使用 AI 而使用 AI。" Token 消耗量，他强调，"在任何意义上都不是影响力的衡量标准。" 但喊话止不住出血。如今 Meta 正在搭建系统性管控机制——产品开发人员和工程师团队已经做出一个名为 AI Gateway 的实时仪表盘，全公司 AI 用量和支出一目了然，异常消费峰值自动触发预警。未来几周管控工具将向更多员工铺开，2027 年前推进结构化的 Token 预算分配。

· · ·

At the same time, Meta is weaning employees off third-party AI tools. The newly formed Applied AI Engineering department has been tasked with upgrading MetaCode (formerly Devmate), the internal coding assistant, to replace Anthropic's Claude — currently the go-to coding tool for Meta engineers. The team is generating programming challenges to produce high-intensity reinforcement learning training data. Meta hasn't locked the door entirely: staff can still use OpenAI, Anthropic, and Google tools. But internal tools get priority. The dual strategy — restrict externally, promote internally — lays bare the tightrope tech giants walk between betting big on AI and keeping costs in check.

与此同时，Meta 正把员工从第三方 AI 工具上"断奶"。新成立的应用 AI 工程部门（Applied AI Engineering）被要求升级内部编程助手 MetaCode（前身 Devmate），目标直指取代 Anthropic 的 Claude——目前 Meta 工程师写代码的首选工具。团队正在生成编程挑战题，为高强度强化学习训练提供数据。Meta 没有彻底封死大门：员工仍能用 OpenAI、Anthropic 和谷歌的工具，但内部工具优先级被明确调高。外部限制、内部推广——这套组合拳暴露了科技巨头在激进押注 AI 和精细化成本管控之间的走钢丝。

引自