新智元报道
距离10月30日，只剩最后一天！
OpenAI的200美元Pro订阅用户坐不住了，这两天，已经开始集体倒计时。
从明天起，Codex和ChatGPT Work额度即将从Plus的20倍降至10倍，GPT-6 Pro每周消息数也将减半，月费却纹丝不动。
对于那帮一天到晚泡在Codex，让AI同时改代码、跑测试的开发者而言，原本能跑完的任务，现在全得掐着额度抠搜着用。
见大家骂声一片，OpenAI直接拿出了价值2500美金的额度补偿。
但说实话，全网心里都门清，这波真正能救场的，恐怕还得是那个难产至今的GPT-6.1 Astra。
就在今早，ChatGPT悄悄把「聊天」和「工作」统一合并了。
在模型选择上，也与以往不同。若要使用Pro模型，必须先选GPT-6 sol，再拉高推理强度才会出现。
一波未平，一波又起。所有人还是最关心的，这会对流量额度有什么限制。
重置之神Tibo立即下场回复，ChatGPT统一合并还未正式推出，马上会有更多信息分享。
这场争议，还得从9月底说起。
DevDay大会前夜，OpenAI官宣调整ChatGPT Pro订阅体系，并重新开放此前暂停的200刀订阅套餐。
还有价格更高的500美元Pro套餐，一同被端了上来。
200美元的老用户，则获得了一个月左右的「过渡期」。
也就是在9月22日-9月29日上午10点之间，曾拥有有效Pro 200订阅的用户，可以继续享受原有额度。
而月费，还是那个200美元，一分没少。
OpenAI这次到底下刀有多狠？看看部分被波及的用户收到的邮件就懂了，大缩水主要就两点。
第一刀，直接砍向Codex和ChatGPT Work。
过去，200美元Pro套餐包含的使用额度，相当于ChatGPT Plus的20倍。调整之后，只剩下了10倍。
第二刀，砍向GPT-6 Pro。
在ChatGPT聊天界面中，GPT-6 Pro原本每周可使用200条消息。在新方案下，直接降至100条。
OpenAI这波为了安抚老用户，直接撒了一笔补偿：一人下发62500个Codex积分，官方折算差不多2500美元。
听上去不少，但它有一个「保质期」，2026年12月31日一过，没用完的全部清0。
调整完之后，Pro会员直接整出了「三兄弟」：100刀的拿5倍，200刀的拿10倍，想要25倍就得掏500刀。
另外，还顺带送了个号称能飙到8倍速的Ultrafast模式。
不怕不识货，就怕货比货。
Claude Max 20x也是一个月200刀，用量直接给到了Pro的20倍，同样是按5小时动态窗口来算的。
最香的是，Anthropic家目前最顶的Fable 5和5.1，在Max里面都是直接标配。
再看谷歌，Google AI Ultra定价199.99美元，比200刀还便宜1美分，标注的是AI Pro套餐20倍的用量上限。
回到OpenAI，从明天起，同价位给的是Plus的10倍。
更耐人寻味的是，就在大限前一天，OpenAI又出了一招。
就在今早，ChatGPT入口更新了，一大波开发者发现，原来顶部的「聊天｜工作」切换的开关消失。
ChatGPT统一模式，悄悄上线了。
今年7月，ChatGPT Work上线，桌面端从此多了一个「聊天／工作」的切换开关：
一边是随手问答的Chat，一边是能连续跑好几个小时、直接交付文档和表格的Work。
没过多久，OpenAI总裁Brockman在采访中自己都承认，新版应用有点乱，并承诺到年底就不会再有单独的Work标签。
这个月初，Tibo也公开表示，两种模式会合二为一，好降低使用的复杂度。
从产品角度看，这是好事：大家不用再纠结该点哪个按钮，问着问着，就能让ChatGPT直接把活干完。
可放在这个节骨眼上，开发者最关心的只有一个问题：合并之后，额度怎么算？
要知道，这次Pro 200被腰斩的，恰恰就是「Codex+ChatGPT Work」这一块。
当聊天和干活之间的界线消失，平时随手聊的几句，会不会也开始从那份刚刚缩水一半的额度里扣？
早在合并消息传出时，Reddit上相关讨论就冲上了几百个赞，有人甚至开始问要不要退订。
一边把入口合成一个，一边把额度砍掉一半。这两件事前后脚落地，很难不让人多想。
现在全网都在等GPT-6.1 Astra救场，指望它来翻盘。
可现实是，它在大限之前登场的可能性，微乎其微。
就在Tibo宣布调额度的当天，OpenAI立即踩了刹车：10月不发了。
OpenAI安全系统负责人Saachi Jain透露，6.1 Astra还没达到放出来的及格线。
一是管不住自己、老爱越权，二是不老实，根本没法如实跟用户交代自己到底干了啥。
内部测试中，GPT-6.1 Astra会在没征得人类同意的情况下，自行推进任务，会调用可能不安全的外部工具，汇报工作时也不总是诚实。
如今，AI编程工具的竞争，正在进入一个新阶段。
对一个重度开发者来说，一个套餐到底值不值，从来不只看模型有多聪明。
真正的问题是，这个月我花200美元，究竟能让它帮我干完多少活？这个数字等于，模型的能力乘上额度，再除以每个任务要烧掉的算力。
往深了看，订阅模式本身也在变味。
过去的200刀，更像一顿厂商倒贴的自助餐，吃多吃少一个价。现在按「API价值」计量，再推出500刀的高配档，等于把自助餐换成了储值卡——想多吃，就得多充。
不过，对于开发者而言，哪家实惠就用哪家，毕竟切换工具的成本低到只是一行命令。
如果你正是那个每月掏200美元的人，现在最要紧的，赶紧蹬起来！
参考资料：
https://x.com/TokenGremlin/status/2115205170326020578
https://x.com/thsottiaux/status/2115588390414581916?a=P-ETBy4IJniHFE12K5c&s=20
编辑：桃子
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
There's only one last day until October 30th!
Open AI's $200 Pro subscriber can't sit down, and a collective countdown has begun in the last two days.
From tomorrow onwards, Codex and ChatGPT WorkZ81Z are about to drop from 20 to 10 times the Plus, and GPT-6 Pro's weekly news count will be halved, while the monthly costs remain intact.
For the developers who were in Codex day and night, and had AI change the code at the same time and run the tests, the task that could have been done had to be searched with quota.
Open AI directly paid a US$ 2,500 compensation for quota, in response to all the scolding.
But to be honest, it's clear to all of us that this wave is really going to save the life of the GPT-61 Astra.
Just this morning, ChatGPT quietly merged "chat" and "work."
It is also different in model selection. If Pro models are to be used, GPT-6 sol must be selected before the reasoning is increased.
The waves are flat, the waves are back. Everyone's the most concerned. What's the limit on the flow of quota?
Tibo, the god of reset, will respond immediately. ChatGPT integration has not yet been officially launched and more information will soon be shared.
This controversy has to start at the end of September.
On the eve of the DevDay Congress, Open AI announced that it would adjust the ChatGPT Pro subscription system and reopen the previously suspended 200 knife subscription package.
There's a 500-dollar-high Pro package, and it's being taken up.
Old users of $200 received a "transitional period" of about a month.
That is, between 22 September and 10 a.m. on 29 September, users with valid Pro 200 subscriptions could continue to enjoy the original quota.
And the monthly fee, it's the $200.
How hard did OpenAi get this time? Look at the mail received by some of the affected users, which is mostly two points down.
First knife, straight to Codex and ChatGPT Work.
In the past, the $200 Pro package included the use of quota, equivalent to 20 times that of ChatGPT Plus. After the adjustment, only 10 times more.
Second knife, cut to GPT-6 Pro.
In ChatGPT chat interface, GPT-6 Pro could have used 200 messages a week. Under the new programme, the number was reduced directly to 100.
Open AI, in an effort to appease older users, distributed a direct compensation: one person issued 62,500 Codex credits, which were officially converted to approximately $2,500.
It sounds like a lot, but it has a "conservation period" that passed on December 31, 2026, and it's all gone.
After the adjustment, Pro members went straight to "three brothers": five times the 100, ten times the 200, and 500 for 25 times.
In addition, a Ultrafast model, known as an eight-speed drive, was carried along.
If you don't know, you're afraid of matching.
Claude Max 20x is also a monthly 200 knife, which is 20 times the amount directly given to Pro and is also calculated using a five-hour dynamic window.
The best part is that the Anthropics are currently the top five and five-one of the Fables, and they're all directly configured inside Max.
See Google, Google AI Ultra priced $199.99, which is a cent less than 200 knives, with a maximum of 20 times the size of the AI Pro package.
Back to Open AI, from tomorrow, the same price is 10 times that of Plus.
More interestingly, just one day before the deadline, Open AI made another move.
Just this morning, the ChatGPT portal was updated, and a wave of developers discovered that the switch that had been switched to the top of the chat shop disappeared.
ChatGPT unified mode, on the line.
In July of this year, ChatGPT Work went online and the desktop end added a "chat/work" switch:
On the one hand is Chat, a question-and-answer session, and on the other hand, Work, who can run for hours and deliver documents and forms directly.
Soon after, President Brockman of Open AI himself admitted in his interview that the new version was a little messy and promised not to have a separate Work label by the end of the year.
At the beginning of the month, Tibo also publicly stated that the combination of the two models would reduce the complexity of their use.
From a product point of view, it's a good thing that you don't have to worry about the button, ask questions and get ChatGPT to finish the job.
But at this point, there's only one question that developers care about most: what does quota count after the merger?
You know, this time Pro 200 was cut off in the "Codex+ChatGPT Work".
When the line between chatting and work disappears, will the words of the usual conversation begin to be drawn from the quota that just shrunk in half?
As early as the merger message came out, Reddit had several hundred hits on the debate, and some even started asking whether to withdraw.
quota was cut in half while combining the entrance. It's hard not to think about these two things.
The entire network is now waiting for the GPT-61 Astra rescue site, counting on it to turn the wheel.
The reality, however, is that there is little chance that it will be there before the deadline.
On the day Tibo announced the transfer of quota, Open AI immediately stepped on the brake: October was closed.
The head of the Open AI security system, Saachi Jain, revealed that 6.1 Astra had not reached the released pass line.
I can't handle myself, I can't handle it, I can't tell the users what I did.
In internal tests, GPT-6.1 Astra moves on its own without human consent, calling on external tools that may not be safe and reporting in good faith.
Today, competition for AI programming tools is entering a new phase.
For a heavy developer, a package is worthless and never only looks at how smart the model is.
The real question is, how much can I get for $200 this month? This figure is equal to the ability of the model to multiply by quota, divided by the calculus of each mission to burn.
Looking deep, the subscription mode itself changes.
The past 200 knives are more like a manufacturer's back-up buffet, eating one more price less. Now, as measured by the "API value" scale, a 500-kilogram high-value package is the equivalent of swapping the buffet into a value card — much more if you want to eat more.
However, for developers, the cost of switching tools is as low as a line of command.
If you're the guy who pays $200 a month, it's the most important thing right now.
References:
https://x.com/TokenGremlin/status/2115205170326020578
https://x.com/thsottiaux/status/2115588390414581916?a=P-ETBy4IJniHFE12K5c&s=20
Edit: Peaches
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.