OpenAI Reportedly Completes 10T+ Parameter “Bel” Pretraining For GPT-6

OpenAI has reportedly completed a massive next-generation pretraining run internally codenamed “Bel,” with more than 10 trillion total parameters, according to an unverified report circulating in the AI community. The claim originated with an August 25 post on X by account @synthwavedd, which described Bel as the successor to an earlier project called “Doug” and said the new pretraining run is expected to serve as the foundation for future systems including Astra and GPT-6. OpenAI has not publicly confirmed Bel, its parameter count or the reported roadmap.

If the report is accurate, Bel would represent a major milestone in OpenAI’s effort to scale its underlying AI models. The reported roadmap places Bel not as a finished consumer product but as a base model requiring additional reinforcement learning and post-training before systems such as Astra or GPT-6 could be released. The source also claims OpenAI believes Bel could potentially support a future model approaching an AGI threshold, although that remains speculation rather than a demonstrated achievement.

OpenAI Bel Pretraining At A Glance

Because OpenAI has not officially acknowledged the project, the following figures should be treated as reported rather than confirmed.

Key Details

Particular	Reported Details
Project codename	Bel
Company	OpenAI
Reported status	Pretraining completed
Reported parameter count	10 trillion+
Previous project	Doug
Reported successors / products	Astra, GPT-6
Training stage	Pretraining
Further work	Reinforcement learning / post-training
Reported role	Future foundation model
AGI connection	Potential future AGI-threshold base
Official OpenAI confirmation	None reported
Original claim	@synthwavedd on X

The original post described Bel as a successor to Doug and said it could become a post-GPT-6 foundation for future frontier models. 

What Is Bel?

Bel is reportedly an internal OpenAI pretraining project rather than a publicly available model.

According to the claim, the system is intended to function as a foundation model that can subsequently be transformed through reinforcement learning and other post-training techniques into more specialized or capable systems.

Doug
  ↓
Bel
  ↓
Large-Scale Pretraining
  ↓
Post-Training + Reinforcement Learning
  ↓
Astra
  +
GPT-6
  ↓
Future Frontier Models

The source further claims that OpenAI expects Bel to remain useful as a foundation even after GPT-6 launches.

That would make Bel more of a model-generation platform than a single consumer-facing product.

Bel Reportedly Exceeds 10 Trillion Parameters

The most eye-catching claim is the reported size of the pretraining run.

According to @synthwavedd, Bel contains more than 10 trillion total parameters.

For context, the report compares its scale with GPT-4.5. However, OpenAI has never publicly disclosed the parameter count of GPT-4.5, meaning the comparison should also be treated cautiously. 

Reported Scale

OpenAI Model Development
        ↓
Doug
        ↓
Bel
        ↓
10T+ Parameters
        ↓
Astra / GPT-6
        ↓
Future Frontier Models

A parameter count alone does not determine model capability. Architecture, training data, compute, inference methods, reinforcement learning and post-training can all materially affect performance.

Why A 10 Trillion-Parameter Model Would Matter

If the reported figure is accurate, Bel would represent an enormous increase in model scale.

A model with trillions of parameters requires massive computing resources for:

• Pretraining
• Data processing
• Interconnects
• Memory
• Storage
• Checkpointing
• Evaluation
• Reinforcement learning
• Inference

The challenge therefore extends far beyond simply building a larger neural network.

Requirements For Frontier-Scale Training

Infrastructure	Why It Matters
GPUs / AI accelerators	Model computation
High-speed networking	Connect training clusters
Memory	Store model states and activations
Data centers	Provide physical infrastructure
Power	Run large-scale compute
Cooling	Manage heat
Storage	Training data and checkpoints
Software stack	Coordinate distributed training

This is one reason the report’s connection to OpenAI’s broader infrastructure expansion is significant.

Bel Is Reportedly The Base For GPT-6

The report says Bel is expected to serve as the foundation for GPT-6, with additional reinforcement learning applied afterward. 

That distinction is important.

Pretraining creates the underlying model, but modern frontier systems typically undergo additional stages designed to improve:

• Instruction following
• Reasoning
• Safety
• Tool use
• Coding
• Agentic behavior
• Reliability

Pretraining
    ↓
Base Model
    ↓
Supervised Fine-Tuning
    ↓
Reinforcement Learning
    ↓
Safety + Alignment
    ↓
Product Model

Therefore, completion of Bel’s alleged pretraining would not mean GPT-6 itself is finished.

Astra Is Another Reported Destination

The same source connects Bel with a future system called Astra.

According to the report, Astra would be derived from Bel and receive additional reinforcement learning.

However, OpenAI has not officially confirmed the relationship between Bel and Astra.

The term Astra has appeared in speculation around OpenAI’s future AI systems, but there is currently no official technical documentation establishing the reported architecture.

Reported OpenAI Roadmap

Stage	Reported Role
Doug	Earlier pretraining project
Bel	10T+ parameter foundation
Astra	Future system based on Bel
GPT-6	Future GPT generation based on Bel
Post-GPT-6 models	Potentially continue using Bel as a foundation
AGI-threshold model	Speculative future possibility

This roadmap comes from the reported source and should not be treated as an official OpenAI product roadmap. 

Bel Is Not Reported To Be An AGI Model

One of the most important distinctions is that Bel itself is not being reported as AGI.

The claim says OpenAI believes the model could potentially serve as the foundation for a future model that reaches an AGI threshold.

That is fundamentally different from saying the pretraining run has achieved AGI.

Bel
 ↓
Foundation Model
 ↓
Further Training
 ↓
More Capable System
 ↓
Potential AGI-Level Capability

The AGI component therefore remains a forward-looking claim.

Why Pretraining Completion Matters

Large-scale pretraining is one of the most difficult stages of developing a frontier AI system.

A successful run indicates that the organization has managed to coordinate:

• Model architecture
• Training infrastructure
• Data pipelines
• Distributed computing
• Optimization
• Fault tolerance
• Checkpointing
• Evaluation

For an alleged 10T+ parameter model, the engineering challenge would be substantially greater than for previous generations.

OpenAI’s Scaling Challenge

The Bel report arrives after a period in which OpenAI has continued releasing new models and model variants but has not publicly confirmed a similarly dramatic new foundation-model pretraining milestone.

The report claims that Bel follows an earlier effort called Doug.

Other reporting has characterized the period after GPT-4o as unusually long for OpenAI’s frontier-model development, although the company’s internal training work remains largely undisclosed. 

This makes the reported completion of Bel particularly significant if it is eventually confirmed.

The Competitive AI Race Is Intensifying

OpenAI’s reported progress comes as competitors continue to scale their own AI systems.

The industry currently includes major model developers such as:

• Anthropic
• Google DeepMind
• Meta
• xAI
• Alibaba
• DeepSeek
• Z.ai

The competition is no longer focused only on benchmark performance.

Companies are also competing on:

• Training scale
• Inference cost
• Coding
• Agentic capabilities
• Multimodal reasoning
• Long-context performance
• AI infrastructure

Frontier AI Competition

Area	Competitive Focus
Model scale	Larger and more capable foundations
Reasoning	Complex problem solving
Agents	Autonomous task completion
Coding	Software development
Multimodal AI	Text, image, audio and video
Infrastructure	Training and inference capacity
Cost	Lower cost per useful task

A successful Bel training run would potentially strengthen OpenAI’s position in several of these categories.

Anthropic Is Reportedly Concerned

The original claim also makes a prediction about Anthropic.

It says OpenAI believes Anthropic does not have a strong public response prepared for Astra and attributes this largely to computing constraints.

The report says Anthropic expects much of the remainder of 2026 to favor OpenAI but believes it could return to the lead in early 2027. 

These are competitive assessments contained in the original report, not statements confirmed by Anthropic.

Reported Competitive Timeline

2026
OpenAI → Reported Bel / Astra Advantage
        ↓
2026 End
OpenAI Expected To Hold Lead
        ↓
Early 2027
Anthropic Reportedly Expects Comeback

This means the source itself does not describe the competition as permanently settled.

Compute Is Becoming The Strategic Battleground

The reported Anthropic response highlights one of the biggest issues in frontier AI: compute availability.

Building increasingly capable AI systems requires enormous amounts of accelerator capacity.

That makes access to:

• AI chips
• Data centers
• Electricity
• Networking
• Cooling
• Capital

increasingly strategic.

More Capable Models
        ↓
More Compute
        ↓
More AI Chips
        ↓
More Data Centers
        ↓
More Power
        ↓
Higher Capital Requirements

The companies that can secure sufficient infrastructure can potentially train larger models faster.

OpenAI’s Stargate Infrastructure Is Relevant

OpenAI’s broader Stargate infrastructure initiative has been designed to expand AI computing capacity.

However, claims linking particular future models to specific Stargate infrastructure should not be treated as confirmed unless OpenAI provides those details.

The current Bel report establishes no official connection between the reported model and a particular data-center cluster.

Parameter Count Is Not Everything

The reported 10T+ parameter figure is impressive, but it should not be used as a direct measure of intelligence.

Two models with different parameter counts can have substantially different capabilities.

Other factors include:

Factors Influencing AI Capability

Factor	Importance
Parameters	Model capacity
Training data	Knowledge and patterns
Data quality	Learning efficiency
Architecture	Computational efficiency
Training compute	Optimization
Reinforcement learning	Reasoning and behavior
Post-training	Usefulness and alignment
Inference techniques	Runtime performance

This is particularly important because OpenAI has not disclosed the architecture or training methodology of Bel.

A Successful Pretraining Run Still Leaves Major Work

Even if Bel’s pretraining has been completed exactly as reported, significant work would remain before any public product could emerge.

OpenAI would need to conduct:

• Post-training
• Reinforcement learning
• Safety testing
• Red teaming
• Benchmark evaluation
• Tool integration
• Product optimization
• Deployment testing

The difference between a base model and a production AI system can therefore be substantial.

Why The Report Should Be Treated Carefully

The central Bel claim comes from a single social-media post.

OpenAI has not issued a public announcement confirming:

• Bel’s existence
• The 10T+ parameter count
• Doug’s relationship to Bel
• Bel’s relationship to GPT-6
• Bel’s relationship to Astra
• The reported AGI objective
• Anthropic’s alleged compute position

Independent coverage of the claim has also emphasized that the information remains unverified. 

Confirmation Status

Claim	Status
Bel exists	Unverified
Pretraining completed	Unverified
10T+ parameters	Unverified
Successor to Doug	Unverified
Foundation for GPT-6	Unverified
Foundation for Astra	Unverified
AGI-threshold potential	Speculative
Anthropic compute constraints	Unverified

This distinction is especially important because the original source is an insider-style social-media claim rather than an OpenAI technical publication.

What Could Confirm The Bel Report?

Several developments could establish whether the claims are accurate.

These could include:

1. An OpenAI technical announcement.
2. A model card.
3. A research paper.
4. A product announcement referencing Bel.
5. Independent evidence from developers or infrastructure partners.
6. Public technical details about GPT-6 or Astra.

Until such evidence appears, the 10T+ parameter figure should remain categorized as a report.

What A 10T+ Model Could Mean For AI

If confirmed, Bel would signal that OpenAI continues to believe scaling large foundation models remains a viable route toward substantially greater AI capability.

That would be significant because the industry has increasingly emphasized efficiency, reasoning and specialized post-training.

A model at this scale could potentially provide a much larger base from which reinforcement learning and agentic training can extract capabilities.

10T+ Parameter Foundation
          ↓
Large Knowledge Representation
          ↓
Reasoning + RL
          ↓
Tool Use + Agents
          ↓
Multimodal Capabilities
          ↓
Future Frontier Systems

The actual benefit, however, would depend on the quality and efficiency of the training process.

The Bigger Picture

The reported completion of OpenAI’s 10T+ parameter Bel pretraining run, if eventually confirmed, would represent a major milestone in the industry’s race to scale frontier AI. The reported model is described as the successor to Doug and a foundation for future systems including Astra and GPT-6. Crucially, however, the claim currently originates from an X post and has not been confirmed by OpenAI through a technical paper, model announcement or official statement. 

The more consequential aspect of the report is the suggestion that OpenAI intends to use Bel beyond a single GPT generation. If the same foundation can support GPT-6, Astra and potentially later systems, OpenAI may be pursuing a strategy in which an enormous pretraining run becomes a reusable platform for multiple rounds of reinforcement learning and post-training. The reported AGI connection should nevertheless be viewed as an ambition or expectation, not evidence that AGI has been achieved. 

Looking Ahead

The next major development to watch is official confirmation. If OpenAI publicly acknowledges Bel or releases GPT-6/Astra-related technical information that matches the reported architecture, the 10T+ parameter claim would become considerably more credible. Until then, the exact scale, training methodology and intended relationship between Bel, Astra and GPT-6 remain unknown. The distinction between completing a pretraining run and delivering a production-ready frontier model is also critical, because substantial reinforcement learning, safety testing and post-training would still be required.

For the broader AI industry, the report highlights how the competition is increasingly becoming a contest over compute, capital and model-training infrastructure as much as algorithms. A verified 10T+ parameter foundation would reinforce the idea that OpenAI is continuing to pursue aggressive scaling even as rivals focus on efficiency and lower-cost models. Whether larger pretraining runs ultimately translate into a decisive capability advantage will depend not simply on parameter count, but on how effectively OpenAI converts that enormous foundation into reasoning, agents and reliable real-world performance