OpenAI is rumored to release GPT-6.2 Astra this week; Claude Code product lead Cat Wu has defected…

GPT-6.2 Astra is coming, Lina.

According to the buzz The Information released in the early hours of this morning, as soon as this Friday local time, OpenAI's next-generation model family is about to launch a group assault, rolling out the all-new Astra, Sol, and a new model focused on cybersecurity all at once.

And before that, netizens across the internet had already found plenty of clues about the new models.

In the early hours of yesterday, a model-tracking bot suddenly caught two wild codenames in OpenAI's API list: rc-ada-or-vm and rc-ada-sh-vm.

Although the official side deleted them in seconds, 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 were suspected to have already been quietly deployed on Google Cloud.

Originally, everyone thought going from 6.0 to 6.2 was just a routine iteration.

Until X netizen imjustnewatai leaked that the internal model to be released this week is extremely expensive, but its intelligence is absurdly high.

Even if the new model's reasoning tier is set to the lowest "low" setting, its performance can grind the previous generation's full-power "max" tier GPT-6 Astra into the ground.

Even more brutal is the accompanying Sol model.

People used to think lightweight versions were just watered-down goods distilled from large models, but this time Sol went straight into the same devilish training camp as the internal flagship, using the same high-intensity game-theoretic and extreme-environment reinforcement learning, and its capabilities took off on the spot.

And on the front-end aesthetics front, developer Pankaj Kumar has already run an early test.

He casually threw in a prompt asking the model to make a hand-drawn anime-style dynamic website, in which a child flies a red kite, and the kite pulls apart the dark clouds to reveal the sunrise.

The resulting product not only had top-tier lighting and aesthetic sense, but even the subtle details—the wind stirring, birds flapping their wings, the child's tiny movements tugging the kite string—were all arranged clearly and completely.

In addition, X netizen Fix also posted a screenshot SVG from inside Minecraft that was allegedly generated by GPT-6.2 Astra, and then showed the effect of having Astra animate the SVG.

Of course, it is still impossible to confirm whether these works really came from GPT-6.2 Astra, so everyone can treat them as community leaks for now.

If netizens' leaks are still just puzzle pieces, then rewind to November 3, when Altman casually dropped a bombshell from the stage—looking back now, it was purely pre-hype for Friday's blockbuster.

Getting worked up, Altman said they used this upcoming next-generation GPT model to spin up 10,000 Agents at once, running nonstop for four days straight, and directly proved the first Millennium Prize Problem solved by AI...

Stop, stop, stop.

Patting my own wallet, last night I used Codex to run a slightly larger refactoring task, and after a night's sleep the quota bar hit bottom and threw an error. So the whole world's compute has been taken by Altman to do scientific research and backyard steelmaking, huh...

Not only is it grinding problems in the math exam room, the new model has even started crossing over to steal hardware engineers' jobs.

OpenAI recently showed off a handheld hardware experiment called TamaGPT on its official blog. Two employees, with zero manual modifications, relied entirely on Astra to handle everything from the casing's mechanical structure and PCB circuit board routing to the low-level firmware, and in less than a week produced designs that let a contract manufacturer build a physical prototype.

Since February this year, the daily token consumption of OpenAI's hardware team has surged 23-fold, with each person burning 212 million tokens per day on average, and that mysterious internal model's score on the 3D engineering benchmark BenchCAD 2.0 jumped directly from the previous generation's 0.594 to 0.875.

It is worth mentioning that right at the critical moment when OpenAI is racing ahead on both software and hardware fronts, the backyard of neighboring Anthropic suddenly caught fire.

And the trouble was not somewhere else, but precisely Anthropic's most proud flagship product, the Claude Code team.

Within just ten days, four core members staged a collective walkout.

First, on October 24, Thariq Shihipar, a core creator of the CLI and Agent SDK who often interacts intensely with us in the community, suddenly announced his farewell.

On the same day, Anthony Morris, the "father of desktop" who single-handedly expanded Claude Code from a black-box command line into a desktop GUI client, also waved goodbye and left.

Before everyone could recover, three days later on October 27, engineering manager Sunil Shah, who was responsible for project scheduling and engineering processes, also appeared on the departure list.

Until yesterday, the most fatal hammer blow landed.

Cat Wu, the overall product lead for Claude Code and Cowork and the absolute number-two figure on the team, announced her departure on X yesterday, and today she directly flashed her new work badge.

Her next stop is none other than OpenAI!

You should know that Cat Wu and Claude Code's number-one, Boris Cherny, were known in Silicon Valley as the golden duo "Batman and Robin." Back then, it was she who carried Claude Code all the way from an internal reinforcement learning prototype to today's phenomenal official product.

What's more interesting is that last July she and Boris had once job-hopped together to Cursor's parent company Anysphere, but only stayed two weeks before the two hand in hand flashed back to Anthropic, which at the time became a much-told tale in the industry.

But this time, Robin has truly gone solo for good.

And the destination is extremely delicate: she went straight into OpenAI's Core Products department, reporting to none other than the Codex number-one we all know, the cyber godfather Tibo Sottiaux.

Funnily enough, just the day before, another OpenAI big shot, Andrew Ambrosino, suddenly tweeted a Git merge meme, saying that from today on everyone could send him feedback on all Apps, since he personally loves a perfect merge.

As soon as the tweet went out, outsiders speculated that Tibo was about to be laid off. Seeing his "cyber retirement" rumors growing more absurd, Tibo himself had to step in and wryly borrowed Mark Twain's classic line to debunk it.

"Reports of my retirement have been greatly exaggerated."

Good grief, everyone finally put their hearts back in their stomachs. After all, if Tibo really retired, when we spend the weekend furiously grinding code and burn through all our quota, who would we go to in the X comments section to beg for a reset? o(╥﹏╥)o

But then again, word on the street is that Anthropic, in order to withstand the heavy punch of GPT-6.2 Astra, is holding back a big move and preparing to roll out Claude Fable 5.5 around the same time.

But looking at the situation now, Anthropic probably has a huge headache. A year ago, when Cursor poached people, Anthropic could still rely on sentiment to coax "Batman and Robin" back together.

Now, on the eve of the two sides' new models facing off head-to-head, its own "Robin," who best understands how to play the Agent product game, is already sitting in the office of Tibo on the opposite side.

So what else is there to say, Tibo? Since you not only didn't retire but also gained a top general, once the new model drops this week, go ahead and arrange our reset first (doge).

This article comes from the WeChat public account "APPSO," author: APPSO, which discovers tomorrow's products, published by 36Kr with authorization.