在2025年5月到12月28日之间，“氛围编码”从偶尔成功变成了预期：提示词现在常常能直接产出可用代码。产出规模不再主要受写代码限制，而更多受推理时间以及那一小部分需要深度思考的工作限制。他倾向先从CLI开始，让代理能运行命令、验证输出、快速闭环；在此过程中他“烧掉了大量token”，以建立对哪些事该花几分钟或几小时的直觉。

GPT‑5被视为主要解锁点，而从5/5.1到5.2的跃迁被形容为“巨大”。一个月前他做了Oracle：一个能跑GPT‑5 Pro的CLI，可上传文件与提示词并在之后取回会话；Pro能在约50个网站上“速跑”检索，有时约10分钟，有时超过1小时。随着5.2到来，“问oracle”的频率从每天多次降到每周几次。他还强调知识截止期差距：8月底对比3月中旬，约5个月。

与Opus 4.0相比，codex可能会在开始写代码前默读10–15分钟；即便在可比任务上耗时约4倍，他认为总耗时反而更低，因为几乎不需要反复修补“修补的修补”。一次仅两句话的提示词驱动了完整的Zig改写，过程跑了约5小时仍一次性交付可运行结果。他通常同时推进3–8个项目、用队列堆任务，并能并行跑4个重构，每个约1–2小时。他的配置把工具输出上限设为25,000 token，并在233,000 token处自动压缩，以适配约272–273k的上下文窗口。

Between May and Dec 28, 2025, “vibe coding” shifted from occasional success to an expectation: prompts now often yield working code out of the box. Output volume is constrained less by writing and more by inference time plus the small fraction of work that needs hard thinking. He prefers starting as a CLI so agents can run commands, verify output, and close the loop quickly after “burning a lot of tokens” to learn what should take minutes versus hours.

GPT-5 was the main unlock, and the step from 5/5.1 to 5.2 was described as “massive.” A month earlier he built Oracle, a CLI that runs GPT-5 Pro, uploads files plus a prompt, and retrieves sessions later; Pro could speedrun 50 websites, sometimes in 10 minutes, sometimes taking >1 hour. With 5.2, “ask the oracle” fell from multiple times per day to a few times per week. He also highlights a cutoff gap: end-of-August vs mid-March, about 5 months.

Compared with Opus 4.0, codex may silently read code for 10–15 minutes; even when it takes 4× longer, he claims total time drops because fewer fixes-of-fixes are needed. A two-sentence prompt drove a full Zig conversion that ran 5 hours and still landed working. He typically juggles 3–8 projects, queues tasks, and runs 4 refactors at once that each take 1–2 hours.