花旗、Adobe、Atlassian等限制员工使用高成本大模型，企业AI开销飙升
英文摘要
Multiple large enterprises are restricting employee access to high-end large language models to control surging AI costs. Citi blocked Claude Opus 4.6, 4.7, and GPT-5.5 on June 24; Adobe ended unlimited Claude access on June 30; Atlassian saw monthly AI costs jump from $5 million to over $15 million and introduced a cost dashboard. The shift from fixed annual contracts to per-token pricing has caused uncontrolled token consumption, with Accenture reporting massive waste on tasks like PDF-to-presentation conversion. GitHub plans to switch to open-source models and test per-user billing, while Amazon removed a leaderboard that inadvertently encouraged excessive AI use before imposing token caps.
中文摘要
多家大型企业为控制飙升的AI成本，开始限制员工使用高成本大语言模型。花旗于6月24日封禁Claude Opus 4.6、4.7及GPT-5.5；Adobe于6月30日终止Claude无限使用协议；Atlassian月度AI支出从500万美元增至超1500万美元，并上线成本看板。企业AI服务从固定年费转向按量计费，导致词元消耗失控，埃森哲发现大量开销用于PDF转PPT等琐事。GitHub计划改用开源模型并测试按人计费，亚马逊下架鼓励滥用的AI用量排行榜后随即设定词元上限。
关键要点
Citi blocked employee access to Claude Opus 4.6, 4.7, and GPT-5.5 on June 24, initially planned to restore July 1, due to high per-interaction token cost.
花旗银行6月24日封禁员工对Claude Opus 4.6、4.7及GPT-5.5的访问，因单次交互词元消耗过高，原计划7月1日恢复。
Atlassian's monthly AI spending tripled from $5 million to over $15 million, leading the company to eliminate unlimited usage and display cost per employee on a dashboard.
Atlassian月度AI支出从500万美元增至超1500万美元，取消无限使用权，并在看板展示每位员工的AI成本。
Adobe's unlimited Claude access agreement expired on June 30 and was not renewed, effectively limiting the model's use.
Adobe的Claude无限使用协议于6月30日到期且不再续约，限制该模型的使用。
GitHub is moving to open-source models and testing per-user billing to reduce costs, while Amazon removed an internal usage leaderboard and later imposed token limits.
GitHub为降低成本计划改用开源模型并测试按人计费；亚马逊下架内部用量排行榜后对员工设定词元上限。
Accenture observed huge token waste on trivial tasks like converting PDFs to slides and now offers clients 'token cost economics' consulting.
埃森哲发现大量词元消耗用于PDF转PPT等琐事，现向客户提供‘词元成本经济学’咨询服务。