GPT-6.2 and 'Bell': Every OpenAI Leak Explained So Far
Leaked model IDs, a rumored 10-trillion-parameter model called Bell, and benchmark leaks hint at OpenAI's next GPT release.
What is GPT-6.2 “Bell” and why is everyone talking about it?
Bell is the reported codename for a massive new base model built inside OpenAI, referenced online under the internal name “GPT-6.2 Beetlejuice.” Leaks from late October and early November 2025 point to model IDs briefly appearing in OpenAI’s API, internal usage stats showing a model called GPT-6.2 Astra, and an unnamed internal model scoring far above OpenAI’s current public model on an engineering benchmark. None of it is officially confirmed, but the pattern of leaks is unusually consistent.
TL;DR
- A model ID called GPT-6.2 Beetlejuice reportedly surfaced in an internal OpenAI tracker on October 30, with “Bell” as its nickname, pointing to a new base model rather than just a minor update.
- Codex users spotted GPT-6.2 Astra in their usage breakdowns on October 31, even though the app still displayed GPT-6 Astra as the active model, suggesting silent testing on real users.
- Two unlisted model names, RC-Ader-VM and RC-Aer-SH-VM, briefly appeared in OpenAI’s official API model list on November 4 before vanishing, with “RC” widely read as “release candidate.”
- The Information reported new OpenAI models could ship as soon as November 6, potentially including updated Astra, Soul, and Cyber Focus models, while cautioning that timelines can shift.
- An unnamed internal model reportedly scored 87.5% on a hard engineering benchmark called BenchCAD 2.0, compared with GPT-6 Astra’s best score of 59.4%, according to a chart circulated from an OpenAI blog post.
- Reported parameter counts for Bell range from 10 to 12 trillion, which would make it far larger than most current open models, though the exact figure is unverified.
- Sam Altman reportedly said the AI behind OpenAI’s Navier-Stokes proof will become the next GPT, tying a widely publicized math result directly to this unreleased model.
Auditable. Scalable.
What has actually leaked so far?
The timeline starts on October 30, when an account on X claimed a model ID called “GPT-6.2 Beetlejuice” had appeared in an internal OpenAI tracker, with “Bell” floated as its nickname. Tech outlet WCCF Tech reported a similar screenshot and said the model would likely sit on top of Bell, described as a large new base model under development inside the company.
The next day, people combing through Codex usage data (Codex is OpenAI’s coding agent) noticed a name they hadn’t seen before: GPT-6.2 Astra. The app still displayed GPT-6 Astra as the active model, but the usage breakdown told a different story. For at least one user, the unlisted model accounted for close to 20% of their usage over a week, which reads like quiet A/B testing with real users who weren’t told what they were using.
Around the same time, Thibault Sautier, who leads Codex at OpenAI, was mid-way through a public 28-day run of daily improvements for Codex users. Near the end of that stretch, Codex announced a reset of banked usage limits, something many read as a signal to save usage headroom for something bigger. Sautier then posted two cryptic phrases on X days apart: “GPT fat” and “chat BLT.” Both got picked up quickly by AI-watcher accounts as hints about a large new model, with “fat” read as a nod to Bell’s size.
On November 4, the Chinese tech outlet 36Kr reported that two previously unseen model names, RC-Ader-VM and RC-Aer-SH-VM, appeared briefly in OpenAI’s official API model list before disappearing within minutes. Developers who tried calling them got capacity errors rather than a “model not found” response, suggesting the models were real but not yet ready for traffic. “RC” is commonly understood as “release candidate,” the build that typically precedes a final release.
That same day, reporters Stephanie Palazzolo and Tiffany Lee at The Information, a publication with a strong track record on OpenAI reporting, said new models could arrive as soon as November 6, potentially covering updated Astra, Soul, and Cyber Focus model lines, while noting that such timelines often slip. Shortly after, a separate account posted what appeared to be a cloud hosting listing referencing “GPT-6.2 Astra” and “GPT-6.2 Soul,” unconfirmed but consistent with the Information’s reporting.
What makes Bell different from a routine update?
The leaks point to two separate things. One is GPT-6.2 Astra, which looks like an improved version of OpenAI’s existing public model. The other is Bell, described in multiple reports as an entirely new base model, not an incremental patch.
A base model is the raw, pre-trained system an AI company builds before adding chat behavior, reasoning training, or safety tuning. It’s trained on enormous amounts of text and other data in a process called pre-training, the most computationally expensive stage of building a large language model. Everything added afterward, instruction-following, reasoning, safety guardrails, is built on top of that foundation. A stronger base model theoretically raises the ceiling for everything built on top of it.
- Name
- Maria Alvarez
- Date of birth
- SSN
- Diagnosis
- Balance due
- $412.80
Reported parameter counts for Bell vary. WCCF Tech put the figure at over 10 trillion parameters after pre-training finished. 36Kr’s reporting cited a range of 10 to 12 trillion. For comparison, many of today’s leading open-weight models sit around 1 trillion parameters, so if these figures hold up, Bell would be among the largest models ever trained. More parameters generally mean more capacity to store patterns and knowledge, though raw parameter count alone doesn’t guarantee better real-world performance.
Some more extreme claims are circulating too, including a reported benchmark score on the Artificial Analysis Intelligence Index and specific token pricing, but these numbers lack verified sourcing and should be treated as rumor rather than fact. The Beetlejuice codename and the Bell nickname are the most consistently corroborated parts of this story. The exact size, pricing, and benchmark scores are not.
Why would a bigger base model matter now?
For the past couple of years, a common argument in AI research has been that scaling up pre-training alone had stopped producing big gains, and that most recent progress came from teaching models to reason for longer before answering (so-called “test-time compute” or chain-of-thought reasoning). If OpenAI has in fact built a dramatically better 10-trillion-plus parameter base model, that would suggest the older scaling approach still has headroom. Combined with better reasoning techniques and more capable agent systems, those three trends compounding together would represent a faster pace of improvement than any one alone.
What do the early benchmark numbers show?
The most concrete data point so far doesn’t come from a leak but from OpenAI’s own blog post, titled “Accelerating Engineering R&D in the Physical World.” It introduces BenchCAD 2.0, a benchmark testing whether an AI can perform real engineering design work: writing code that reconstructs 3D parts and assemblies, and tracing how components on a circuit board connect to each other. It’s a demanding test because it requires understanding geometry, physical measurements, and spatial relationships, not just text.
A chart from that post, shared by AI commentator accounts and summarized in 36Kr’s reporting, showed two models. GPT-6 Astra, OpenAI’s current best public model, topped out at 59.4% even at its highest reasoning setting. An unnamed internal model scored 75.8% at its lowest effort setting, already beating Astra’s best result by 16 points, and climbed to 87.5% at its highest setting, a 28-point gap over Astra. OpenAI has not confirmed that this internal model is Bell or GPT-6.2, but the timing lines up closely with everything else that has leaked.
What has Sam Altman actually said about the next model?
On November 2, speaking at the ADIPEC energy conference in Abu Dhabi, Altman reportedly told attendees that the AI system behind OpenAI’s Navier-Stokes result is the next GPT and that it would be released soon, according to accounts from people present.
Without shadow IT.
The Navier-Stokes equations describe fluid motion and are central to physics and engineering. Parts of the underlying mathematics had resisted formal proof for roughly two centuries, making it one of the seven Millennium Prize Problems, each carrying a $1 million prize from the Clay Mathematics Institute. In September, OpenAI said an internal model, more capable than GPT-6 Astra, produced a solution along with a machine-checkable formal proof. The effort reportedly involved around 10,000 AI agents working in parallel over several days without human direction, processing roughly 130 billion tokens, a volume OpenAI researcher Noam Brown compared in an interview to thousands of years of continuous human thinking compressed into about 88 hours. The Clay Mathematics Institute said the problem appears to be resolved, though full verification will take time, and there’s an ongoing credit dispute with an NYU mathematician who published related work the same day.
Altman has since described the moment as one of the most nerve-wracking of his career, separate from his public ADIPEC comments, and claimed the Navier-Stokes effort cost around $10 million but could be repeated for a fraction of that cost within a year. He’s also made broader claims, including that AI now exceeds top human scientists in some respects, a framing he tends to use often and that’s worth treating with some skepticism on its own.
Is “GPT-6.2” likely to be a minor update?
Version numbers like “.2” suggest small, incremental change, but recent history argues against reading it that way. Anthropic’s Claude Opus 5.5, released September 22, looked like a modest point release but delivered large jumps on real benchmarks: Terminal-Bench 4 scores rose from 52.3% to 66.4%, and GDPval, a benchmark built from real job tasks, jumped by more than 100 points. Anthropic also reported it caught more known bugs at its lowest effort setting than Opus 5 did at its highest, while cutting per-token costs by around 20%. It quickly became the default model across Claude’s products. That precedent suggests a “.2” label from OpenAI could mask a similarly large capability jump rather than a routine refresh.
Frequently Asked Questions
What is GPT-6.2 Beetlejuice?
It’s a model ID reportedly spotted briefly in an internal OpenAI tracker in late October 2025. “Beetlejuice” appears to be an internal project name, and “Bell” is the nickname attached to what’s described as a large new base model associated with it.
How big is the Bell model supposed to be?
Reports vary, with figures ranging from just over 10 trillion to 12 trillion parameters. These numbers come from unverified leaks and tech outlet reporting, not from an official OpenAI statement, so they should be treated as unconfirmed.
What is a base model and why does it matter?
A base model is the raw, pre-trained system an AI company builds before adding instruction-following, reasoning, and safety training on top. A stronger base model generally raises the ceiling for everything built on it afterward, which is why a new, much larger base model would matter more than a routine update.
Is GPT-6.2 confirmed to be releasing soon?
No official release date has been confirmed by OpenAI. The Information reported that new models, potentially including updated Astra, Soul, and Cyber Focus versions, could arrive within days of early November 2025, but cautioned that such timelines often change.
Did OpenAI really use 10,000 AI agents to solve a math problem?
OpenAI said an internal model contributed to work on the Navier-Stokes equations, and reporting around the project described roughly 10,000 AI agents collaborating over several days. The Clay Mathematics Institute said the result appears to settle the problem, though full verification is ongoing, and a dispute over credit with another researcher remains unresolved.