GPT-6.2 Astra is coming, Li Na

GPT-6.2 Astra is coming, Li Na.

According to the leaks released by The Information in the early hours of today, as soon as this Friday local time, OpenAI's next-generation model family will be launched in a surprise drop, rolling out the brand new Astra, Sol, and a new model focused on cybersecurity all at once.

Prior to this, netizens who follow tech news around the world have already found plenty of clues about the new models.

Yesterday in the early morning, the model tracking bot suddenly spotted two wild codenames in OpenAI's API list, rc-ada-or-vm and rc-ada-sh-vm.

Although the official deleted the post immediately, it was already too late.

By the early hours of today, netizen lyra even dug up that GPT-6.2 Astra and 6.2 Sol are suspected to have been quietly deployed on Google Cloud.

We originally thought that the upgrade from 6.0 to 6.2 was just a regular iteration.

Until X user imjustnewatai broke the news that the internal model to be released this week is extremely expensive, but its intelligence level is incredibly high.

Even if you adjust the inference gear of the new model to the lowest low level, its performance can completely outperform the previous full-powered max version of GPT-6 Astra.

Even more impressive is the supporting Sol model.

In the past, people always thought that the lightweight version was a downsized product distilled from the large model, but this time Sol directly entered the same "devil training camp" as the internal flagship, using the same high-intensity game theory and extreme environment reinforcement learning, making its capabilities soar to a new level.

As for the front-end aesthetic performance, developer Pankaj Kumar has already conducted a closed beta test in advance.

He casually entered a prompt, asking the model to create a dynamic website in hand-drawn anime style, where a child is flying a red kite, and the kite pulls apart dark clouds to reveal the sunrise.

The finished product not only has excellent lighting and aesthetic design, but also perfectly presents all the subtle details, including the wind blowing the grass, birds flapping their wings, and the child holding the kite string tightly.

In addition, X user Fix also posted a screenshot SVG from the game Minecraft, which is allegedly generated by GPT-6.2 Astra, and then showed the effect of Astra animating the SVG file.

Of course, it is still impossible to confirm whether these works are truly generated by GPT-6.2 Astra, and you can treat them as community rumors for now.

If the netizens' revelations are just scattered puzzle pieces, then going back to November 3, the explosive news that Sam Altman casually mentioned on the stage of his speech, is now clearly a warm-up for this Friday's big release.

When he got excited, Sam Altman said that using this upcoming next-generation GPT model, they launched 10,000 Agents at once, ran nonstop for 4 days, and directly proved the first Millennium Prize problem solved by AI...

Wait a minute.

Touching my own wallet, last night I used Codex to run a slightly larger refactoring task, and when I woke up after sleeping, my quota was completely exhausted and an error popped up. It turns out that all the computing power in the world was taken by Sam Altman for scientific research, right...

Not only solving problems in the mathematical field, the new model has even started to take jobs from hardware engineers across industries.

OpenAI posted a palm-sized hardware experiment called TamaGPT on its official blog a while ago. Two employees did not make any manual modifications, and Astra took charge of all designs from the shell mechanical structure, PCB circuit board routing to the underlying firmware. Within less than a week, the design drawings were completed and the foundry produced a physical prototype.

Since February this year, the daily Token consumption of OpenAI's hardware team has skyrocketed by 23 times, with each person burning 212 million Tokens per day. The score of that mysterious internal model on the 3D engineering benchmark BenchCAD 2.0 has directly soared from 0.594 of the previous generation to 0.875.

It is worth mentioning that at the critical moment when OpenAI is advancing rapidly on both software and hardware tracks, a sudden problem broke out at neighboring Anthropic.

And the problem did not occur elsewhere, but in the Claude Code team, the ace product that Anthropic is most proud of.

In just ten days, four core backbones collectively left the company.

First, on October 24, Thariq Shihipar, the core creator of CLI and Agent SDK who often interacts closely with the community, suddenly announced his departure.

On the same day, Anthony Morris, known as the "Father of Desktop Client" who single-handedly expanded Claude Code from the black box command line to a desktop GUI client, also left the company.

Before everyone could recover, three days later on October 27, Sunil Shah, the engineering manager in charge of project scheduling and engineering processes, also appeared on the resignation list.

Until yesterday, the most shocking news came.

Cat Wu, the overall product lead of Claude Code and Cowork, the absolute second-in-command of the team, announced her resignation on X yesterday, and even showed her new work ID today.

Her new employer is exactly OpenAI!

You know, Cat Wu and Boris Cherny, the top leader of Claude Code, are known as the golden duo "Batman and Robin" in Silicon Valley. It was her who led Claude Code all the way from an internal reinforcement learning prototype to the phenomenal official product it is today.

What's more interesting is that last July, she and Boris once jumped to Anysphere, the parent company of Cursor, together. But they only stayed for two weeks, and the two returned to Anthropic hand in hand, which became a well-known story in the industry at that time.

This time, Robin has completely left to start a new journey.

And her new position is very intriguing: she directly joined OpenAI's Core Products department, and her reporting line goes to the well-known top leader of Codex, the "cyber godfather" Tibo Sottiaux.

Interestingly, the day before, another OpenAI executive Andrew Ambrosino suddenly posted a tweet playing on a Git merge meme, saying that from now on everyone can send all App feedback to him, because he loves to make a perfect merge.

After the tweet was posted, the public widely speculated that Tibo was about to step down. Seeing that the rumors of his "cyber retirement" became more and more absurd, Tibo himself had to come forward to refute the rumor, using Mark Twain's famous classic quote.

"The reports of my retirement are greatly exaggerated."

Great, everyone finally breathed a sigh of relief. After all, if Tibo really retired, when we burn through all our coding quota over the weekend, who can we go to in the X comment section to ask for a reset o(╥﹏╥)o.

That said, there are widespread rumors in the industry that Anthropic is preparing to launch Claude Fable 5.5 at the same time to counter the heavy blow from GPT-6.2 Astra.

But looking at the current situation, Anthropic is probably having a huge headache. A year ago when Cursor poached their employees, Anthropic could still use shared memories to coax both "Batman and Robin" back.

Now, on the eve of the head-to-head showdown between the two new models, their "Robin" who knows best about Agent product strategies has already sat in Tibo's office on the opposite side.

So what else can we say, Tibo? Since you are not retiring and have gained a top talent, when the new model is released this week, remember to arrange the quota reset for us first (doge).

This article is from WeChat Official Account "APPSO", the author is APPSO that discovers tomorrow's products, 36Kr publishes it with authorization.