OpenAI API briefly shows dual virtual machine model codename rc-ada, withdrawn after a few minutes

OpenAI's next ace has already been prepared.

Yesterday, Codex lead Tibo dropped a cryptic hint, "Le GPT Fat," and the whole internet has been wildly speculating that GPT-6.2 Astra is about to arrive.

As luck would have it, today two never-before-publicized mysterious model IDs suddenly flashed onto OpenAI's official API model list—

rc-ada-or-vm

rc-ada-sh-vm

Just minutes later, OpenAI swiftly pulled these two models from the API list, and every developer quick enough to fire off a call ran straight into a "503 capacity limited error."

Also today, OpenAI's official blog published a long hardware engineering article, quietly revealing a not-yet-released "mysterious internal new model."

Its entry-level run score alone surpassed GPT-6 Astra's highest full-power tier, with its peak surging to 87.5%.

At almost the same time, netizens posted a wave of GPT-6.2 Astra tests, instantly maxing out the internet's expectations.

This long-sequence demo of "a pelican riding a bicycle by the seaside" is packed with detail and stunning in texture, with rendering quality that can actually rival Fable 5.5.

Not only that, GPT-6.2 Astra's hand-coded pure front-end animation is also ready to "consume."

In the image below, it completed an animation from 0 to 1, but in terms of camera design and detail granularity, it still got a harsh lesson from Fable 5.5.

The same animation made by the big shot Token Gremlin was done by GPT-6.2 Astra itself.

The classic version of "Minecraft" first produced an SVG image, then generated the animation.

In the SVG test, four major models simultaneously drew a PS5 controller, and GPT-6.2 Astra and Fable 5.5 are simply the new SOTA.

Listed for a few minutes then swiftly withdrawn

Dual-VM architecture and the GPT-6.2 twin stars surface

This late-night catch was simultaneously spotted by developers Pankaj Kumar, Pradeep Kapoor, and the automated tracking sentinel AI Tracker Bot.

Following OpenAI's usual release rhythm, 24 to 48 hours before a new model is officially announced, it often undergoes a brief mounting test on the production environment API gateway.

In the currently leaked dual-model codenames, the prefix rc is short for Release Candidate, indicating the model has completed internal training and testing and can be pushed online at any time; ada is the internal codename for this generation of models.

What really hides the mystery is the two vm (virtual machine) suffixes.

or-vm handles task decomposition and global scheduling, equivalent to the "brain virtual machine" that issues orders; sh-vm is the "hands and feet virtual machine" that runs command lines, writes code, and launches software inside an isolated sandbox.

English text is being prepared.

Another group of developers pointed the clues toward two new models that had just leaked from Codex internal logs: gpt-6.2-astra and the codename "Betelgeuse" gpt-6.2-betelgeuse (Bel for short).

The former has quietly taken over nearly 20% of Codex's internal traffic, while Bel, named after the red supergiant, is precisely the ultra-large mother model with a parameter scale of 100 trillion to 120 trillion, exactly matching the "fat GPT" Tibo mentioned.

Also, just two days ago at the Abu Dhabi summit, Altman personally admitted that the AI behind mobilizing 10,000 Agents to work continuously for 70 hours and single-handedly crack the millennium problem "NS equations" was precisely the "next-generation GPT" about to be unveiled.

The two candidate versions that flashed onto the API this time are very likely the GPT-6.2 "twin stars" about to be listed.

Mysterious new model "crushes Astra's full-power version at entry level"

If the API flash only revealed a name, then the blog post "Accelerating Engineering R&D in the Physical World," published by OpenAI at almost the same minute, laid bare the terrifying muscle of this generation of new models.

The sharp-eyed evaluation agency LuminaBench immediately zeroed in on the first benchmark chart in the blog post.

In this benchmark chart named BenchCAD 2.0, OpenAI rarely placed its flagship GPT-6 Astra and a mysterious opponent signed "internal model" in the same table.

This leaderboard tests all hardcore engineering work: 100 real cases requiring the model to write code while invoking physical engineering knowledge to rebuild 3D parts and assemblies, and also sort out the connection relationships on PCB circuit boards.

Digging into the chart data underlying OpenAI's webpage, the gap between the two generations of models is quite stark.

The current flagship GPT-6 Astra goes from the lowest tier all the way up to the fifth tier, and even pushing inference compute to the limit, its score stops at only 59.4%.

But that mysterious internal new model, right from the start, took 75.8% at its lowest tier, 16 percentage points higher than Astra's full-power limit, while the per-question cost was actually cheaper; continuing to add inference compute, the score soared all the way to 87.5%, raising the ceiling by 28.1 percentage points.

Haozhe Zhang, first author of the BenchCAD 2.0 benchmark, exclaimed in amazement, saying outright that the leap from 0.594 to 0.875 was too shocking.

Netizen Prinz did the math based on Astra's API pricing: this mysterious internal model can get an entry-level high score for just $11 per question, and after $37 of deep thinking it can deliver a quasi-industrial-grade answer.

Seen this way, the logic instantly clicks.

To automatically build 3D assemblies, check tolerances, and run PCB routing in BenchCAD 2.0, the model must be able to call CAD scripts and simulation software by itself inside a sandbox, which exactly corresponds to the dual-virtual-machine architecture of rc-ada-or-vm and rc-ada-sh-vm that flashed in the API.

From Tibo's teaser "Le GPT Fat," to the rc-ada dual-virtual-machine models that flashed in the API backend and were withdrawn in seconds, to the 87.5% internal model whose cards were revealed in the same minute on the official blog, all the puzzle pieces are now in place.

Since the Release Candidate (RC) has already been mounted on the API gateway, only the final countdown remains before OpenAI officially drops its big move.

Reference materials:

This article comes from the WeChat public account "Xinzhiyuan," author: ASI Revelation, editors: Moses and Taozi, published by 36Kr with authorization.