OpenAI leaks next-gen GPT-6.2 candidates with dual-VM architecture

If you were watching OpenAI's developer portal late at night, you might have caught a rare glimpse of the next big thing.

While the internet was still gossiping about when the next frontier model would drop, two unannounced model IDs quietly popped up on OpenAI's official API list: rc-ada-or-vm and rc-ada-sh-vm. Within minutes, OpenAI yanked them back down, leaving developers who tried to ping them facing 503 capacity errors.

Even so, those brief minutes gave away the whole game plan.

Dual virtual machines built for agents

The leaked names tell a very specific technical story.

The rc prefix stands for Release Candidate, signaling the model has already cleared internal evaluations. But the real giveaway is the vm suffix attached to both models.

According to technical breakdowns circulating among developers, or-vm acts as an orchestrator, handling global planning and task decomposition. Meanwhile, sh-vm runs inside an isolated sandbox, executing bash commands, writing code, and driving software tools.

Baking this dual-VM setup straight into the API foundation means OpenAI isn't just shipping another chatbot. It is shipping an engine purpose-built for autonomous, multi-step agent workflows.

This also connects the dots with Codex lead Tibo's recent teaser hinting at "Le GPT Fat." Internal logs spotted by developers indicate that alongside gpt-6.2-astra, an enormous base model nicknamed Betelgeuse (Bel) packing an estimated 10 to 12 trillion parameters has been quietly taking shape.

A mystery model obliterates the benchmark

Right around the same time the API leaked, OpenAI published an engineering post titled Accelerating Physical World Engineering, quietly showing off the actual muscle behind these models.

In the post, a benchmark called BenchCAD 2.0 pitted the current flagship, GPT-6 Astra, against an unnamed "internal model" across 100 complex physical engineering challenges, including 3D assembly and PCB board routing.

The gap was staggering.

GPT-6 Astra hit a ceiling of 59.4% even when pushed to its highest inference tier. The mystery internal model scored 75.8% at its lowest baseline setting—easily beating Astra's absolute best—and rocketed to 87.5% at maximum compute. That lifts the benchmark ceiling by 28.1 percentage points.

BenchCAD 2.0 lead author Haozhe Zhang expressed sheer disbelief at the leap. Outside estimates suggest the model could deliver near-industrial CAD work for around $37 in deep-thinking compute per problem.

To build assemblies and verify tolerances in BenchCAD, the model has to pilot CAD scripts and simulation suites in a live environment, which mirrors the dual-VM leak perfectly.

Why it matters

We think this leak marks a fundamental pivot in how OpenAI wants us to interact with frontier intelligence.

Until now, running reliable agents required developers to glue together external frameworks, separate task planners, and clunky local sandboxes. By packaging the orchestrator and the execution environment together directly at the API layer, OpenAI is turning multi-step agency into a plug-and-play utility.

For engineers and builders, crossing the 87.5% threshold on BenchCAD means AI is moving from casual prototyping to serious, industrial-grade mechanical design. Spending $37 on machine reasoning to solve a complex engineering headache is already dramatically cheaper than hours of senior engineering time.

For the wider industry, this shifts the frontier war away from synthetic coding benchmarks and straight into the messiness of real-world software tooling. If your model cannot operate inside a sandbox with reliable execution loops, you are about to fall behind.

The catch, of course, is safety and control. Giving dual-VM setups deeper autonomous reach inside execution sandboxes raises the stakes for drift and runaway loops. But since release candidates are already touching production gateways, we will find out very soon how OpenAI plans to keep them in check.