视频 · 前往原文观看我们感兴趣的是，AI 模型的改进如何改变开发者的工作方式。具体而言，开发者在多大程度上能更高效地完成他们已经在做的任务，又在多大程度上，更好的模型能够让他们完成以前无法触及的工作？
为了回答这个问题，我们与芝加哥大学布斯商学院的 Suproteem Sarkar 教授合作，研究了 2025 年 7 月至 2026 年 3 月期间，使用 Cursor 的 500 家公司的开发者工作习惯。这八个月的时间窗口内，包含了 Opus 4.5 和 GPT-5.2 的发布，这两个模型在 AI 编程能力上带来了阶跃式的进步。
我们的论文发现，更好的 AI 会带来更大的 AI 需求。这与类似杰文斯悖论（Jevons-like effect）的现象一致，即效率的提升反而增加了总消耗量，而非减少。在研究期间，AI 使用量（定义为每位用户每周平均发送的消息数）增加了 44%。
这种增长并非立竿见影或均匀分布。我们观察到，开发者首先利用更好的模型来完成更多复杂度相似的工作，之后才开始承担更复杂的任务。此外，这种转变尤其集中在金融、媒体和广告等行业，这些领域的竞争压力和全新机遇可能推动了其采用。
媒体、软件和金融行业领先
在我们研究的每个行业中，使用量都有所增加，但某些行业的增长幅度更大。特别是，媒体和广告行业的增长最为显著，每位用户发送的消息数增加了 54%，其次是软件和开发者工具行业（+47%），以及金融和金融科技行业（+45%）。
我们推测，在金融领域，更好的 AI 可能引发一种军备竞赛式的动态：一旦某家公司利用 AI 获得了交易优势，其他公司就会面临跟随的竞争压力。而在媒体和广告领域，其机制可能有所不同：能力更强的模型拓展了全新的机遇，企业则纷纷加以利用。
复杂度的右移
起初，开发者们利用改进后的 AI 模型做着与之前类似的工作，但在经过 4 到 6 周的滞后之后，我们观察到他们开始将模型用于更复杂的任务。总体而言，在研究期间，“低复杂度”消息的数量增加了 22%，而“高复杂度”消息的数量增长了 68%，其中大部分增长发生在最后六周。
在论文中，我们假设这种延迟既反映了开发者发现更好模型能力所需的时间，也反映了企业需要围绕新能力重新调整其工作流程。
不断变化的任务分布
随着 AI 在代码生成方面的改进，开发者的工作转向管理这些输出。这一变化在我们的数据中清晰可见，我们可以衡量不同任务类别中使用情况的演变。增长最大的领域是文档（+62%）、架构（+52%）、代码审查（+51%）和学习（+50%），而像 UI/样式这类更独立的任务增长则要小得多（+15%）。
这表明，随着 AI 生成的代码扩大了代码库的规模，记录、理解和审查这些代码的需求也相应增长。更大、变化更快的代码库也增加了管理所有部分如何协同工作的复杂性，这或许可以解释架构和部署等跨系统任务的急剧增长。能力更强的模型也可能使开发者更愿意使用智能体来处理这些跨系统任务。
不断扩展的经济活动
围绕 AI 采用的一个核心问题是，它仅仅是促进了现有工作，还是也开辟了新的生产机会。我们的研究表明，它两者兼具，但扩展最终可能成为更重要的故事。
持续改进我们的智能体框架
Stefan & Jediah
利用 warp decode 实现更好的 MoE 模型推理
Less, Federico & Zhiyuan
通过实时强化学习改进 Composer
Jacob, Ben, Nathan & Wanqi
视频 · 前往原文观看We are interested in understanding how improvements in AI models change how developers work. In particular, to what extent do developers perform more of the tasks they were already doing, and to what extent do better models enable work that was out of reach before?
To answer that question, we partnered with Professor Suproteem Sarkar from the University of Chicago Booth School of Business to study the work habits of developers at 500 companies using Cursor, from July 2025 through March 2026. This eight-month window included the releases of Opus 4.5 and GPT-5.2, two models that delivered step-change advances in AI coding capability.
Our paper finds that better AI leads to greater AI demand. This is consistent with a Jevons-like effect, where gains in efficiency increase total consumption rather than reducing it. AI usage, defined as average weekly messages per user, increased 44% during the study period.
The increase wasn’t immediate or uniform. We observed that developers first used better models to do more work of similar complexity, and only later began taking on more complex tasks. Moreover, the shift was especially concentrated in industries like finance, media, and advertising, where competitive forces and greenfield opportunities may have spurred adoption.
Media, software, and finance lead the way
Usage increased in every sector we studied, but the gains were larger in some industries than others. In particular, media and advertising saw the biggest jump, with a 54% increase in messages per user, followed by software and developer tools (+47%) and finance and fintech (+45%).
We hypothesize that in finance, better AI can create an arms-race dynamic, where once one firm uses AI to gain a trading edge, others face competitive pressure to follow. In media and advertising, the mechanism may be different, with more capable models expanding greenfield opportunities that firms take advantage of.
A shift right in complexity
Initially, developers did more of the same with the improved AI models, but after a lag of 4–6 weeks, we observed that they began using models for more complex tasks. Overall, the number of “low complexity” messages increased 22% over the study period, while the number of “high complexity” messages grew 68%, with most of that growth occurring during the last six weeks.
In the paper, we hypothesize that the delay reflects both the time it takes developers to discover what a better model can do, and the need for firms to reorient their workflows around new capabilities.
A changing task distribution
As AI improves at code generation, the developer’s job shifts to managing that output. This change shows up clearly in our data, where we can measure how usage evolves across task categories. The largest increases were in documentation (+62%), architecture (+52%), code review (+51%), and learning (+50%), while more self-contained tasks like UI/styling grew far less (+15%).
This indicates that as AI-generated code expands codebase size, the need to document, understand, and review that code grows in proportion. Larger and faster-moving codebases also increase the complexity of managing how it all fits together, which may explain the sharp growth in cross-system tasks like architecture and deployment. More capable models may also make developers more willing to use agents for these cross-system tasks.
Expanding economic activity
A central question around AI adoption is whether it merely facilitates existing work, or also opens up new productive opportunities. Our study indicates that it does both, but that expansion may eventually be the bigger story.
Continually improving our agent harness
Stefan & Jediah
Better MoE model inference with warp decode
Less, Federico & Zhiyuan
Improving Composer through real-time RL
Jacob, Ben, Nathan & Wanqi