近 1.8 万次真实会话研究:纯 Vibe Coding 更贵更慢,漏洞约四倍
一项基于近 1.8 万次公开 GitHub 编程智能体会话的研究显示,Agent 包办超 99% 代码的 vibe coding 更贵、更慢,漏洞约达人工代码的四倍。
研究者收集了近 1.8 万次来自公开 GitHub 仓库的真实编程智能体会话,覆盖 Claude Code、Codex、Cursor、Gemini CLI 等,并分成三类:人类写全部代码、人机共写,以及 Agent 写出超过 99% 代码的 vibe coding。
结果显示,vibe coding 占比已从 2 月不足 20% 升至近期超过 50%。按最终进入提交的每 100 行计:vibe coding 最贵(约 2.71 美元,人机共写约 0.92 美元)、最慢(约 10.7 分钟对 5.5 分钟),安全方面漏洞约为人工代码的近 4 倍。更意外的是,人机共写不仅优于 vibe coding,也比全程手写更便宜、更快。
会话结构也解释了原因:即便在 vibe coding 中,用户约一半提示在纠正 Agent,而 Agent 主动澄清问题的轮次约仅 3%。样本偏向愿意记录会话的早期使用者,不能代表全体开发者,但它提供了少有的「真实用法」而非纯基准视角。
对团队流程的启示很直接:把 Agent 当结对对象并坚持审阅,往往比「全交给它」或「全自己敲」更划算;流程上应强制代码评审与安全扫描,而不是只看生成速度。
综合来看,读者可先核验官方渠道与发布时间,再结合自身业务边界与合规要求决定是否跟进试点;本文仅整理公开信息,不构成采购、投资或法律建议,关键数字与能力边界请以一手文档为准。
Study of 18k Agent Sessions: Pure Vibe Coding Costs More and Is Less Secure
My bio literally says I'm a "professional" vibe coder. So when a new study measured what vibe coding actually costs, I had to read it. Mostly out of self-defense.
Researchers collected almost 18,000 real coding agent sessions from public GitHub repos. Mostly Claude Code, but also Codex, Cursor, Gemini CLI and others. Real devs, real projects, not a benchmark.
They split them into three types. The human writes everything. The human and the agent write it together. Or vibe coding, where the agent writes over 99% of the code.
Vibe coding went from under 20% of sessions in February to over 50% recently. So at least I'm not alone.
Now the bad part. Per 100 lines that actually made it into a commit, vibe coding was: The most expensive: $2.71, vs $0.92 when working together The slowest: 10.7 minutes, vs 5.5 together The least secure: almost 4x the vulnerabilities of human-written code
And here's what actually surprised me. Working together wasn't just better than vibe coding. It was also cheaper and faster than writing everything by hand.
So the fastest way to code right now isn't the AI doing everything, and it isn't you doing everything. It's the annoying middle, where you actually read what it wrote.
The other number explains why. Users pushed back on the agent in about half of their prompts, even in vibe coding sessions. "That's wrong." "No, the other file." Meanwhile the agent asked a clarifying question in about 3% of its turns.
So half the conversation is you correcting it, and it almost never asks before guessing. So half the conversation is you correcting it, and it almost never asks before guessing.
Turns out vibe coding is just code review where you skip the reading and pay extra.
To be fair, these are early adopters who chose to log their sessions, so it's not everyone. But it's one of the first looks at how people actually use these agents, not how they score on a benchmark.
I'm keeping the "professional" in my bio though. I'm just going to start earning it.
So be honest. Vibe, together, or still typing it all yourself?