Abstract
OpenAI's next flagship model family appears close to release. On 2026-11-04, two never-published model IDs suddenly appeared in the OpenAI official API model list: rc-ada-or-vm and rc-ada-sh-vm. Within minutes, OpenAI removed both, and developers who issued calls in time hit 503 capacity-limited errors. On the same day, an OpenAI engineering blog surfaced a mysterious unreleased internal model whose lowest-tier score already exceeded the full-strength ceiling of GPT-6 Astra, with a peak of 87.5%. This article assembles the available signals: the leaked test demos, the dual-VM architecture implied by the model IDs, the two model names from Codex logs, the NS-equation milestone claimed at the Abu Dhabi summit, and the BenchCAD 2.0 benchmark that placed the mystery model against the current flagship. All facts below are as reported by third-party sources; they are signals, not official confirmations.
1. The Signal That Broke First
The day before, the Codex lead Tibo dropped a code phrase: "Le GPT Fat." The community immediately read it as a hint that GPT-6.2 Astra was imminent. The prediction landed quickly. The next day, two model IDs that had never been public appeared in the OpenAI official API model list for a short window: rc-ada-or-vm and rc-ada-sh-vm. Minutes later, OpenAI pulled both models from the list. Developers fast enough to fire a call ran into a 503 capacity-limited error.
The same day, OpenAI published a long hardware-engineering post on its official blog. Buried inside was a model that had not been announced. Its starting score already crossed the highest full-power tier of GPT-6 Astra, and it climbed to a peak of 87.5%. OpenAI's next-generation card, held back for months, had started to show.
2. The Leaked Test Demos
Almost at the same time, users shared a round of GPT-6.2 Astra tests that raised expectations across the board. A long-sequence demo showed a coastal pelican riding a bicycle. The details were dense and the rendering quality reportedly rivaled Fable 5.5.
GPT-6.2 Astra also produced a pure front-end animation written by hand in code, and it was usable. In the demo it completed the full animation from zero to finished, though in shot design and detail granularity it was still outclassed by Fable 5.5. The well-known account Token Gremlin posted a similar animation that had been made by GPT-6.2 Astra itself. In a classic Minecraft-style example, the model first produced an SVG image, then generated an animation from it. In an SVG test, four large models drew a PS5 controller at the same time. GPT-6.2 Astra and Fable 5.5 were clearly the new state of the art.
3. The Dual-VM Architecture
The late-night packet capture was caught simultaneously by developers Pankaj Kumar, Pradeep Kapoor, and the automated tracking post AI Tracker Bot. The timing follows OpenAI's usual release rhythm: before an official announcement, a new model is often mounted on the production API gateway for a short load test, typically 24 to 48 hours ahead. The leak of these two model IDs fits that pattern.
Reading the IDs, the prefix rc stands for Release Candidate. That label means the model has completed internal training and testing and is ready to ship at any moment. The ada token is the internal codename for this generation. The real clue sits in the two vm suffixes, which stand for virtual machines.
The first, or-vm, handles task decomposition and global scheduling. It acts as the orchestrating brain. The second, sh-vm, runs command lines, writes code, and opens software inside an isolated sandbox. It is the hands-and-feet layer. Winding a dual-virtual-machine environment into the bottom of the API model indicates that this generation is designed for complex, long-horizon agent tasks.
4. Two Model Names from Codex Logs
A separate set of developers pointed to two new models that leaked from internal Codex logs: gpt-6.2-astra and a model code-named gpt-6.2-betelgeuse, abbreviated Bel. The first has quietly taken over close to 20 percent of internal Codex traffic. The second carries the name of a red supergiant star. Its parameter scale is reported at 10 trillion to 12 trillion, making it a very large mother model. That scale lines up directly with the "fat GPT" hinted at by Tibo.
Two days earlier, at the Abu Dhabi summit, Altman acknowledged in person that the AI behind a reported milestone was the upcoming next-generation GPT: 10,000 agents worked continuously for 70 hours and solved a millennium problem, the Navier-Stokes equations. Put together, the two candidate versions that flashed on the API are very likely the GPT-6.2 twin models prepared for release.
5. The Mystery Model's Starting Tier Outruns the Full-Power Astra
If the API flash only exposed a name, the blog post titled on accelerating physical-world engineering research revealed the model's capability. The evaluation organization LuminaBench was first to notice the first scoring chart in the post. The chart, named BenchCAD 2.0, unusually placed OpenAI's flagship GPT-6 Astra and an unlabeled internal model on the same table.
The benchmark tests hard engineering work: 100 real cases. A model must write code, apply physics and engineering knowledge, rebuild 3D parts and assemblies, and untangle the connection relationships on PCB circuit boards.
The underlying chart data shows a wide gap between the two generations. The current flagship GPT-6 Astra climbs from the lowest tier through the fifth tier. Even with inference compute pushed to its limit, its score stops at 59.4 percent. The mystery internal model starts at 75.8 percent on the lowest tier, about 16 points higher than Astra's full-power ceiling, and its per-question cost is cheaper. As inference compute increases, its score climbs to 87.5 percent, lifting the ceiling by 28.1 points.
Haozhe Zhang, the first author of the BenchCAD 2.0 benchmark, was openly surprised, noting that the leap from 0.594 to 0.875 was striking. A user named Prinze calculated the cost using Astra's API pricing: the mystery model returns a high starting score for about 1 dollar per question, and after deep reasoning at around 37 dollars, it can produce an industrial-grade answer.
6. Connecting the Pieces
The logic lines up. To automatically generate 3D assemblies, check tolerances, and run PCB routing inside BenchCAD 2.0, a model must align its own CAD scripts and simulation software with the dual-VM architecture implied by rc-ada-or-vm and rc-ada-sh-vm on the API.
From the "Le GPT Fat" hint, to the rc-ada dual-VM models flashing on the API, to the internal model scoring 87.5 percent over comparable models, and to the BenchCAD 2.0 benchmark backed by millions of hours of training, all the pieces appear to fit. With release candidates already mounted on the API gateway, the countdown to an official announcement is short.
7. What the Signals Suggest About the Product Direction
Reading the architecture clues carefully, several conclusions are reasonable. First, the emphasis on a dual-VM split between an orchestrating brain and an isolated execution sandbox signals a strong bet on long-running agent tasks. The brain VM decomposes work and schedules it; the hands VM executes commands in a sandbox. This separation is a practical design: global reasoning stays in one place, while risky or resource-heavy actions run in isolation.
Second, the scale gap between the Astra-tier model and the Bel mother model matches a common pattern. A large mother model sets the capability ceiling, and a distilled, faster model serves most production traffic. The claim that gpt-6.2-astra already handles close to 20 percent of internal Codex traffic suggests it is being validated in real workflows before the public launch.
Third, the benchmark framing is strategic. BenchCAD 2.0 is not a text-generation test. It requires CAD script generation, engineering knowledge, and assembly reasoning. That positions the model for industrial and physical-world tasks, not only chat or code. The claimed cost efficiency at high tier, around 37 dollars for an industrial-grade answer, points at a pricing model built around task value rather than raw token count.
8. The Gateway as the Release Mechanism
The incident also shows how a frontier model actually reaches production. The 24-to-48-hour pre-announcement mount test on the API gateway is a deliberate step. It verifies that routing, load balancing, and capacity behave correctly under real traffic before the official switch. The 503 errors that hit early callers are a normal side effect of that verification window, not a defect.
For teams that build on frontier models, the release mechanism matters as much as the model. When a new generation appears, the routing layer decides who gets the new model first, how requests fail over, and how usage is measured. A control plane that centralizes this logic, such as an API gateway, lets a team adopt rc-ada-or-vm and rc-ada-sh-vm without rewriting application code. A gateway like 4sapi records which model served each call, applies quotas and fallbacks, and keeps token accounting consistent. That layer turns a model release from a code change into a configuration change. The same mechanism that OpenAI uses internally for its short mount tests is the mechanism teams need externally to consume new models safely.
9. The NS-Equation Milestone and Its Weight
The claim that 10,000 agents ran for 70 hours to work on the Navier-Stokes equations carries its own significance, independent of the model name. Navier-Stokes is one of the millennium problems, and its regularity remains open. A reported large-scale agent effort on it is less about a single answer and more about a style of work: massive parallel agents operating over a long horizon. That style is exactly what the dual-VM architecture is built for. It does not prove the problem was solved, but it does signal that OpenAI is positioning the next model as capable of sustained, coordinated, compute-heavy research.
10. Conclusion
The available evidence points in one direction. OpenAI appears to be preparing a GPT-6.2 family with a split architecture: an orchestrator VM and an execution VM. The release-candidate models have already touched the API. A higher-capability internal model, likely the Bel mother model, appears to set the new ceiling on BenchCAD 2.0, while a faster Astra-tier model absorbs production traffic. The benchmark data, the dual-VM IDs, the Codex log leak, and the summit remark all align.
None of this is an official confirmation. Model availability, pricing, and names are subject to change before announcement. What is structurally important is that frontier releases now depend on the surrounding infrastructure: the routing layer, the capacity test window, and the accounting system. The model is the headline, but the gateway that mounts it, meters it, and fails it over is what makes the release real for developers.
