随着AI编码代理在企业中的普及,按token计费的模式导致开发成本飙升 1。数据显示,单个AI编码任务消耗的token数量是基础聊天提示词的10到50倍 1。这一趋势已产生实际财务影响,据报道Uber在2026年4月就耗尽了其全年的AI预算,原因是扩大了Claude Code的使用 1。市场研究机构Gartner预测,到2028年,AI编码的token成本将超过开发人员的平均薪资 1。此外,用户行为呈现极端分化,前5%的重度用户产生了大部分消耗,最重度用户的用量是最轻度用户的约450倍;如果重度用户从小部分员工增长到半数员工,在人数不变的情况下,月度支出可能增加约4.4倍 1。
针对上述挑战,作者建议企业从单纯限制用量转向优化效率,提倡采用混合模式:将高难度任务路由至前沿API,而将常规工作负载分配给本地部署的开源权重模型 1。该观点文章由HPE制作,并非MIT Technology Review编辑团队撰写 1。文章指出,企业应通过衡量“每个合并拉取请求的成本”(即总AI辅助开发成本除以通过审查并合并的变更数量)而非token消耗来评估效能 1。拥有自有算力基础设施有助于企业在控制敏感代码的同时降低成本 1。以HPE的案例为例,通过将150多个内部应用的常规工作负载路由到本地运行的开源权重模型,token成本降低了30%到50% 1。
As the adoption of AI coding agents accelerates, traditional billing models based on token consumption are driving significant spikes in development expenses 1. To address this, enterprises are advised to shift from merely restricting usage to optimizing efficiency by measuring "cost per merged pull request" 1. This metric calculates total AI-assisted development costs against the number of changes that successfully pass review and integration 1. The approach emphasizes evaluating spending based on tangible shipping outcomes rather than raw data processing volume 1.
The financial pressure on organizations is substantial. A single AI coding task consumes 10 to 50 times more tokens than a standard chat prompt 1. Gartner predicts that by 2028, the token costs associated with AI coding will surpass the average salary of developers 1. Usage disparities exacerbate these issues; the top 5% of heavy users generate most of the consumption, with the most intensive users producing approximately 450 times the volume of the lightest users 1. If heavy users grow from a small minority to half the workforce without an increase in headcount, monthly expenditures could rise by nearly 4.4 times 1. Real-world examples illustrate the severity of budget overruns, such as Uber reportedly exhausting its entire annual AI budget by April 2026 after scaling up its use of Claude Code 1.
To mitigate rising costs while maintaining security, industry experts recommend a hybrid architecture that routes high-difficulty tasks to frontier APIs while assigning routine workloads to locally deployed open-weight models 1. Hewlett Packard Enterprise (HPE) demonstrates the potential of this strategy, reporting a reduction in token costs by 30% to 50% after directing regular workloads across more than 150 internal applications to local models 1. By leveraging proprietary computational infrastructure, companies can control sensitive code access and significantly lower expenses compared to relying solely on third-party cloud services 1. It is noted that this perspective was produced by HPE rather than the editorial team at MIT Technology Review 1.
评论
还没有评论,欢迎留下第一条。