OpenAI's GPT-Next Is Not What You Think (And Codex Has 28 Days)

Author: Universe of AI · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: OpenAI’s GPT-Next Is Not What You Think (And Codex Has 28 Days) · Channel: Universe of AI · 08:13 · Watch on YouTube

Topics: AI agents, AI developer tools, Markets and economy, AI models

Summary

The video discusses OpenAI’s 28-day push to improve Codex amid strong competition from Anthropic’s models. Its first update increases the default Codex speed for GPT-6 Astra and GPT-6.1 by 50%, though the speaker questions whether this will match Anthropic’s speeds or affect token usage. The video also covers speculation about OpenAI’s internal Bell model and a leaked, apparently inexpensive model called GPT Next, which is not presented as a major leap in capability. The speaker views the speed update as a modest start and expects further changes during the remaining 27 days.

Key points

The speaker says Anthropic’s Opus 5.5 and Sonnet 5.5 are matching or outperforming OpenAI models at lower cost.
A Codex leader announced a 28-day period to deliver a broadly useful improvement or reset limits.
OpenAI’s first announced change is a 50% increase in default Codex speed for GPT-6 Astra and GPT-6.1.
The speed improvement is described as a backend change that should not require users to change anything.
The speaker hopes faster generation will not mean the models consume more tokens for the same task.
Bell is described as an internal codename for a new base for future OpenAI models, though its release is only speculation.
A leaked model called GPT Next appears to be inexpensive and capable of coherent outputs, but not a major intelligence upgrade.
The speaker says Anthropic models remain faster in the comparisons discussed and currently have the lead in intelligence.
The video argues that improving Codex and other agent harnesses can improve the overall experience of using models.

Chapters

Start	Chapter	Summary
00:00	Anthropic’s competition and OpenAI’s 28-day push	The video contrasts Anthropic’s recent models with OpenAI’s offerings and introduces the 28-day Codex improvement commitment.
01:00	Codex receives a speed boost	The first update raises default Codex speed by 50%, and the speaker considers its usability benefits and possible token-cost implications.
03:34	Speculation about OpenAI’s Bell model	The speaker discusses Bell as a possible new model base that could help OpenAI respond to Anthropic’s releases.
05:07	What the GPT Next leak suggests	The leaked GPT Next outputs look competent and inexpensive, but the speaker does not see them as evidence of a major new model.
06:41	Comparing speeds and assessing day one	The speaker compares reported model speeds, calls the first update a modest start, and looks ahead to the remaining 27 days.

Quotes

over the next 28 days they will either ship one thing that is a clear improvement and relevant for most Codex/work users or ship a full reset. — 00:30

they have optimized the default speed for Codex to be 50% faster — 01:30

the code name for this model is Bell. — 03:34

This is probably not going to be a model that’s based off of Bell — 06:10

we still have 27 days to go — 08:13

Concepts and entities

OpenAI, Anthropic, Codex, GPT-Next, Bell, GPT-6 Astra, GPT-6.1, Opus 5.5, Sonnet 5.5

Full transcript (original language)

00:00 Over the last few weeks the competition in AI has been intense, especially between Anthropic and OpenAI. Anthropic shipped new models. We got Opus 5.5, Sonnet 5.5, and both of these models surprised a lot of us. Sonnet 5.5 basically was a cheaper version of Opus 5.5, but was producing capable results like Opus 5.5. And if you compare that to anything that OpenAI has at the moment, both of these models are either outperforming or matching OpenAI’s level

00:30 of performance at a cheaper cost. And that has definitely put pressure on OpenAI, and we can notice that because on X, Tibo, the head at Codex and at ChatGPT, who’s actually the one who’s known for resetting everyone’s limit, made a post yesterday basically saying that over the next 28 days they will either ship one thing that is a clear improvement and relevant for most Codex/work users or ship a full reset. And then he said, “Let the improvements begin.”

01:00 Hey, before we get into the video, I just want to take a moment to tell you guys that AI is not slowing down anytime soon. And if you guys want to stay on top of everything that is happening in AI, be on top of all the new model drops, then make sure you guys are subscribed to the channel. Your support means a lot, and now let’s get back into the video. And it looks like they are making this their lock-in time. Only things being worked are simplifications, more efficiency for more usage, groundbreaking features, or new models. So, new models are not out of the, I

01:30 guess, possibility over the next 28 days. Then he said that sometimes you have to invest ahead of the curve, but feedback is clear that you all want things to get simpler on it. And it looks like today, the first day of this, I guess, lock-in mode, we are getting a new update, which is that they have optimized the default speed for Codex to be 50% faster across GPT-6 Astra and GPT-6.1. So, through their subscriptions across all their products and partners using sign-in with ChatGPT. So, this was

02:01 a feature that they actually dropped on Dev Day where if you used to, you know, just use your ChatGPT subscriptions, you kind of have more access now when you use sign-in with ChatGPT. You have access to things like open code, Pi amp, Devin, and more. And he’s saying that none of this is going to affect, you know, anything on our end, but this is going to be a back-end improvement. Now, this back-end improvement is pretty good because currently we are getting GPT 6. Astra and 6.1 Soul at about a rate of 30 token per second, but now after this update once it’s live, it’s going to

02:32 reach 50 token per second, which might not seem like a big deal when you are comparing this to model drops or releases, but we got to remember all of these models we evaluate them based off how easy they are to use and how convenient they are to use. So, if you have a pretty capable model because Astra and Soul are not like horrible models, they’re still strong and capable models, but now those models are faster and nothing changes on your end, you’re going to be happy. Of course, the only risk with this is does this actually mean that we are going to be paying a

03:03 little bit more? Hopefully, the speed of increase in tokens per second does not increase the amount of tokens the models actually going to use to generate the same task at hand. Because if that is the case, then this is counterintuitive. You’re not really getting an update, you’re just getting the model using more tokens per second rather than just a speed improvement if that makes sense. But I don’t think that is going to be the case. I think this is more about just making the models faster on our end, which is going to be a good usability features. Now, what’s interesting is that while all this is happening, people are speculating that

03:34 over the next 28 days, we’re not only just going to to get Codex updates. There’s a possibility that OpenAI drops in new beast model. Now, to understand what people are referencing here, OpenAI has internally been working on a model and the code name for this model is Bell. Now, what this Bell model is going to be, it’s going to be a completely new base for future Open AI models, and this is going to be, I believe, a direct response to Anthropic’s new models, Opus 5.5 and Sonic 5.5, but also direct

04:05 response to Fable 5.5. I feel currently at the moment Anthropic still has another model in the works, and we’ve known that. I’ve I made videos on that, as well. Where Fable 5.1 is being redirected to a stronger version of Fable, which is the 5.5 model. And that model, the outputs that we have seen clearly look like something that we haven’t seen before. The way it’s able to create things so quickly and so fast, it’s not anything that’s out there at the moment. So, Anthropic is definitely, you know, gearing towards dropping that soon, because they want IPO. At the same

04:36 time, Open AI has to drop a model that is going to compete with Opus 5.5, Sonic 5.5, be efficient, be cheap, and not only that, at the same time get ready to compete with Fable 5.5. And that will definitely require the internal model Bell to be released, because no way, as we have already seen, they’re going to be able to drop a stronger model with the code name Duck, which all of the Astra models at the moment or sole models are based off of. So, it looks like it is possible that over the next 27 days, we get a new model from them.

05:07 Before we continue, we just launched the Universe of AI newsletter. If you want to stay on top of AI news without having to hunt for it, link is in the description. Don’t miss out. Now, what’s funny is that while all this is happening, internally there has been a leak that there is a model called GPT next, and this model seems to be a cheap model of Open AI’s. This does not look like a model that, you know, is going to compete with Opus 5.5, for example. I think this is most likely going to be 6.1 Luna, as this person mentions,

05:38 because this just looks like a really cheap model, and it does not look anything drastic, like the outputs that we’re getting from them are not anything crazy, where you would expect it to be from a new code base like Bell. So, this person also mentioned that here are some other leaks you can see that is producing pretty decent results. It’s like not horrible. If we look at this room example that we see, the famous kitchen and everything looks coherent. There’s nothing crazy. Sometimes we have seen like floating dishes or things like that, but that is taken care of by this model. Even if we look at this, I guess

06:10 a voxel garden area or whatever it is with Pokemons in there. I think this is quite amazing to see that the model is able to produce this. And as the person mentioned, this is a cheap model. So, this once again looks like an efficiency play by Open AI. This is probably not going to be a model that’s based off of Bell, but rather just an efficient version of current models. And some people are saying that this could be 6.1 Luna. But whatever the case is, as this person says, this is not going to be the answer to Anthropic’s. This is probably producing outputs like GPT 6.1. So, we

06:41 see that over here. There are no notable improvements expected. The only difference might be in the price, which is the area that Open AI keeps kind of pushing forward, which is good for a lot of users. But if you’re just comparing intelligence of Open AI’s or Anthropic’s model, at the moment Anthropic has the lead. But one thing to understand that today’s is day one. Obviously, this is not the end of the 28-day challenge. In day one, we got a 50% speed boost for GPT 6 Astra and 6.1 Soul. But now if you

07:12 compare this to Opus 5.5 speed, I think it’s still not there. Cuz if you compare the speeds at the moment, you can see that Astra fast is sitting at 75. Opus 5.5 actually sits at 113. Fable 5.1 is actually slower at 85. And then if you have the subscription, Opus 5.5 is at 97 and 96. So, even at the API prices and even at the subscription prices, we are still getting faster models from Anthropic, which I guess is a good thing, but I don’t know why people are

07:42 upset with the update with today. So, this person is not actually as impressed with day one as you mentioned. That’s great. We’re now about halfway to Opus 5.5 speed. I think we deserve a reset. I don’t know how I feel about it. I’m just happy that, you know, the models are getting better and the Codex platform is getting better on the back end. I think making the platform stronger is going to be really important because obviously we use many of these models in those agent harnesses like Codex cloud code. So, if those harnesses technically get better,

08:13 your experience with the model is going to be get better. So, I’m pretty okay with day one. I’m not obviously super excited, but we still have 27 days to go and I’ll be definitely covering every single update. But, that’s it for today’s video. Make sure you guys are subscribed to the channel. [music] Follow our new newsletter as well at universityofai.beehive.com. As well, subscribe to the main channel World of AI and support us on X by following the University of AI Z as well. Until then, I’ll see you guys in the next video.