🔶 最近编码代理的每周配额为何频频随机重置？

一张让我喊“yay”的次数远少于喊“god damn it”的图片。

订阅制编程代理如 Claude Code 和 Codex 因对 LLM 使用设有 5 小时和每周配额而闻名。这两者都可以理解：5 小时配额有助于错峰使用以避免服务器过载，每周重置则可防止用户在一天内耗尽整月用量，这既能防止过载，也避免用户用完就取消订阅。Anthropic 和 OpenAI 都调整过配额限制，从限时翻倍到甚至移除 5 小时配额。

这些模型提供商可以重置所有用户的每周配额，通常是在出现技术故障时作为补偿赠送。例如，如果你使用每月 100 美元的 Codex 计划，在完全用尽配额的前提下，一次每周重置相当于送你 25 美元——如今，随着 Fable 5 和 GPT-5.6 Sol 等顶级 LLM 的成本高昂，这更容易办到。然而，这些配额重置并没有预告，一般也不通过公司官方渠道公布：除非你关注特定人物，比如 Codex 工程负责人 Thibault Sottiaux (Tibo)，否则你只是查看每周配额时才发现它回到了 100%。你甚至可能在真正的重置时间几小时前就获得一次重置，却完全没能受益。

最近，在 Fable 5/GPT-5.6 发布之后，Anthropic 和 OpenAI 对其产品的每周配额重置变得频繁得多。过去两周，OpenAI 直接重置 Codex 每周配额就达 六次：7 月 9 日、7 月 10 日、7 月 10 日（再次）、7 月 14 日、7 月 15 日 和 7 月 17 日。这还没提到很少被讨论的 Codex 银行重置系统：OpenAI 在 7 月 12 日 和 7 月 13 日 赠送了可随时手动使用但 30 天内过期的配额重置。

https://codex-resets.com 跟踪 Codex 重置。

没人想成为那种抱怨免费东西的怪人，因为如果配额重置停了，自己就会成为被埋怨的对象。尽管如此，我还是要当那个抱怨免费东西的怪人。

作为一个只要能轻易避免就不想浪费钱的人，我尽量用光每周配额。我之前用的是每月 20 美元的 Codex 计划，但由于 GPT-5.6 即将到来，加上我频繁撞上 5 小时限制，于是升级到了每月 100 美元的计划。利用 GPT-5.5 的 5 倍提示容量，我构建并测试了不少野心勃勃的项目，不过那就是另一篇博文的事了。我能在一整周内稳定耗尽每周配额，但得打开一个浏览器窗口不停监控 Codex 用量。我甚至在手机上为 5 小时或每周配额重置的确切时间设了提醒，以便能继续提交提示——顺带一提，我真希望有个官方平台端点可以让我用程序检查 Codex 用量和配额重置时间，这样我就能用 vibe coding 自己做个管理应用。在 GPT-5.6 发布前，我曾想：“我是否应该故意在较弱的 GPT-5.5 上用光配额，赌一把 OpenAI 在迎接 GPT-5.6 时会重置配额？”（我没有耗尽配额，而 OpenAI 确实重置了。）

GPT-5.6 Sol 确实是一款名不虚传的出色模型，我不得不创建新项目才能有借口去测试它的极限。即使一直有提示在运行，配额消耗速度也与 GPT-5.5 差不多。尽管如此，GPT-5.6 发布后几乎每次重置发生时，我的每周配额都还剩 50% 以上，这让我觉得自己“浪费”了 12 美元。还有一点是，配额重置后，每周重置时间会不定，所以我必须提交某个提示才能重新触发计时，即使我正在做别的事。

@emollick / X

在过去两周一连串配额重置之后，本该带来的兴奋变成了烦恼，因为我现在必须紧急构思新点子，赶在配额不可避免地再次重置之前把它用掉。随机奖励按理说会带来多巴胺冲击，我却因为浪费配额的感觉和不得不突然改变计划而陷入多巴胺净亏损。我承认，这或许只是我倦怠了、需要休息的信号。

Tibo 发起投票询问是否重置过多，已获数千回应。

面向所有用户的每周配额重置对这些公司来说想必昂贵得离谱，尽管当你每年已经烧掉数十亿美元的资本支出时，这可能只是个舍入误差。近期配额重置激增很可能并非巧合：七月是 LLM 发布的疯狂月份，不仅有 Fable 5 和 GPT-5.6 Sol 进一步推动前沿模型，还有 Grok 4.5、Muse Spark 1.1 和 Kimi K3，在成本/效用曲线上提供了更多选择。一种愤世嫉俗的看法是，每周配额重置并非出于好玩的偶然惊喜，而是意在阻止重度用户在配额自然耗尽后去尝试那些足够有竞争力的竞品。

我并不认为即使竞争加剧，每周配额重置也会永远持续下去，因为要是配额一直如此频繁地重置，它们就毫无意义了。这反而会让我有动力从每月 100 美元的计划降级回每月 20 美元，以避免浪费配额，而这恐怕不是 OpenAI 想要的结果。

The image that makes me say “yay” less often than “god damn it”.

Subscription-based coding agents such as Claude Code and Codex famously have 5-hour and weekly quotas on their LLM usage. Both of these are understandable: 5-hour quotas help stagger usage so the servers don’t get overloaded, and weekly resets prevent users from dumping an entire month’s worth of usage into a single day which a) also prevents overload and b) stops the user from just unsubscribing after they do so. Both Anthropic and OpenAI have played around with quota limits, from doubling them for a limited time to even removing the 5-hour quota.

These model providers can reset the weekly quota for all users, often gifted as compensation in the event of technical glitches on their end. If, for example, you have a $100/mo Codex plan, then a weekly reset is worth $25 to you assuming you fully consume your quota—nowadays with the cost of top-tier LLMs like Fable 5 and GPT-5.6 Sol, that’s easier to do. However, these quota resets are not telegraphed and are generally not announced through company-owned channels: unless you follow specific people such as Thibault Sottiaux (Tibo), the engineering lead for Codex, you just look at your weekly quota and see it’s at 100%. You can even get a quota reset hours before your actual reset and not benefit from it all.

Recently, in the wake of the release of Fable 5/GPT-5.6, Anthropic and OpenAI have been doing weekly quota resets for their harnesses far more frequently. In the past two weeks, OpenAI has directly reset the Codex weekly quota six times : July 9, July 10, July 10 (again), July 14, July 15, and July 17. That’s not even getting into the rarely discussed banked reset system for Codex, where OpenAI gave quota resets July 12 and July 13 which can be manually used at any time but expire within 30 days.

https://codex-resets.com tracks Codex resets.

No one wants to be the weirdo who complains about literally getting free stuff because if the quota resets stop, they will be the one blamed for it. Despite that, I’m going to be the weirdo who complains about literally getting free stuff.

As a person who doesn’t like wasting money if I can easily avoid it, I try to use as much of my weekly quota as I can. I was on the $20/mo Codex plan but with the promise of GPT-5.6 on the horizon and with the frequency I kept hitting the 5-hour limits, I upgraded to the $100/mo plan. By using GPT-5.5, I used the 5x prompt capacity to build and test a number of ambitious projects, but that’s a topic for another blog post. I was able to consistently exhaust my weekly quota over the full week, although I had to have a browser window open with my Codex usage to constantly monitor it. I’ve set reminders on my phone for the exact time a 5-hour or weekly quota resets so I can keep running more prompts—as an aside, I wish there was a canonical platform endpoint to programmatically check Codex usage amount and the quota reset times so I could just vibecode an app to manage for me. Before the release of GPT-5.6, I thought “should I deliberately exhaust my quota all on the weaker GPT-5.5 and gamble that OpenAI does a quota reset to greet GPT-5.6?” (I did not exhaust my quota and OpenAI did indeed do a quota reset)

GPT-5.6 Sol is indeed a great model that does live up to the hype, and I’ve had to create new projects just to have an excuse to test its limits. The rate of quota usage is about the same as GPT-5.5 even with a prompt running at all times. Despite that, almost every time a reset has happened after the GPT-5.6 release, it has been when my weekly quota has been at 50+%, which makes me feel like I “wasted” $12. There’s also the factor that when the quota resets occur, the weekly reset time is unset, so I have to input some prompt just to trigger it again even if I’m doing something else.

@emollick / X

After the flurry of quota resets over the past two weeks, the intended excitement has instead become annoyance as I now have to urgently create new ideas to spin down the quota before it inevitably resets again. Random rewards are supposed to give a dopamine hit but I end up with a net dopamine deficit from both the sense of wasted quota and having to abruptly change my plans. I admit that this may just be a sign I’m burnt out and need to take a break.

Tibo polls if there are too many resets, with a few thousand responses.

Resets of the weekly quota for all users must be ludicrously expensive for these companies, although when you already spend billions of dollars in CapEx per year it’s likely a rounding error. The recent surge of quota resets is likely not a coincidence: July has been an insane month in LLM releases, with not just Fable 5 and GPT-5.6 Sol pushing frontier models even further, but also Grok 4.5, Muse Spark 1.1, and Kimi K3 offering more options across the cost/utility curve. The cynical take is that weekly quota resets are not intended to be fun serendipity, but instead intended to prevent power users from experimenting with sufficiently competitive competitors once the quota naturally runs out.

I don’t expect weekly quota resets to last forever even if competition intensifies, because if quotas keep resetting this frequently they won’t matter at all. It would instead give me an incentive to downgrade from the $100/mo plan back to the $20/mo plan to avoid wasting quota, which I don’t think is OpenAI’s intended goal.