OpenAI has Reportedly Finished Training “Bel” Model with 10 Trillion Parameters
The 10-trillion-parameter claim is unverified, but Stargate’s training advantage makes the underlying AI compute race worth taking seriously.
Last updated on
AI Summary
OpenAI reportedly completed pretraining for Bel, an unconfirmed model exceeding 10 trillion total parameters that may underpin Astra, GPT-6, or a later system and expose a near-term compute gap with Anthropic.
OpenAI has not announced Bel or Doug, disclosed their architecture, or confirmed Astra as GPT-6 as of August 26, 2026. The parameter claim may reflect a sparse mixture-of-experts design and does not establish capability. Astra, however, is acknowledged: OpenAI says it advanced ten mathematics and theoretical computer science problems and triggered stricter cybersecurity controls. Project Stargate has surpassed its initial goal of securing 10 gigawatts by 2029, while GPT-5.5 was trained at its Abilene site using Nvidia GB200 systems.
Anthropic has substantial capacity planned, including more than 220,000 Nvidia GPUs, up to five gigawatts from Amazon, and five gigawatts from Google and Broadcom. Much arrives in late 2026 or 2027, making a temporary frontier-training constraint plausible but unproven.
OpenAI has reportedly completed pretraining for “Bel,” a massive LLM connected to Astra, GPT-6, and perhaps even a later model approaching the company’s threshold for artificial general intelligence. The most eye-catching claim is its scale: more than 10 trillion total parameters.
The report goes further. It says OpenAI believes Anthropic lacks the compute capacity to prepare a competitive response to Astra for a public launch in 2026, potentially giving OpenAI several months at the AI frontier before Anthropic catches up in early 2027.
None of those details has been confirmed by either company. Yet the broader argument is credible: Project Stargate’s most important result may not be cheaper inference today, but the ability to run larger experiments, train stronger base models, and develop multiple generations of GPT systems simultaneously.
Bel Is Still a Report, Not an OpenAI Announcement
The original report from Synthwave, subsequently covered by Wccftech, describes Bel as the successor to another OpenAI pretraining project called “Doug.” It connects Bel to Astra and GPT-6 after further reinforcement learning, while also describing it as a possible base for a post-GPT-6 or “AGI-threshold” system.
That terminology leaves several questions unresolved. It is unclear whether Bel is one model checkpoint, an entire pretraining run, or a reusable foundation from which OpenAI could create several specialized models. The relationship among Bel, Doug, Astra, and GPT-6 is also ambiguous, especially because Astra already exists internally.
As of August 26, 2026, OpenAI has not publicly announced Bel or Doug, disclosed their architectures, or published a parameter count. It has also not officially said that Astra will be marketed as GPT-6. Bel should therefore be treated as an unverified report, not a confirmed OpenAI product specification.
Ten Trillion Parameters Is Not a Capability Score
A model with more than 10 trillion parameters would sound enormous, but total parameter count cannot tell us how intelligent, fast, or expensive the resulting system would be.
The distinction is particularly important for a possible mixture-of-experts architecture. An MoE model can contain many specialized parameter groups while activating only a small selection for each token. Research such as the demonstrated how sparse activation can push total parameter counts into the trillion-parameter range without using every parameter during each forward pass. The Bel report does not confirm MoE, but its emphasis on “total parameters” leaves that possibility open.
Model size also has to be balanced with training data and compute. The Chinchilla scaling research showed that a smaller model trained on substantially more data could outperform much larger, undertrained systems. Data quality, token count, architecture, optimization, multimodal training, and post-training can all matter more than a headline parameter figure.
The claim that Bel is similar in size to GPT-4.5 is equally difficult to evaluate. OpenAI called GPT-4.5 its largest model and said it represented an expansion of pretraining, but the company never published its parameter count. Any numerical comparison between the two rests on information that has not been independently verified.
Astra Is the Nearer-Term Test
Unlike Bel, Astra is an officially acknowledged OpenAI model. On August 1, OpenAI said an internal version of Astra had resolved or made substantial progress on ten long-standing problems in mathematics and theoretical computer science. The company published manuscripts, reasoning walkthroughs, and Lean certificates for the results, while estimating that the model’s solution-generation tokens would have cost roughly $2,000 at GPT-5.6 Sol API rates.
OpenAI provided another signal on August 7. Its internal evaluations found enough progress in agentic coding and cybersecurity that the company said it could not rule out critical cyber capabilities under its Preparedness Framework. OpenAI consequently tightened security controls and paused Astra-related activities that did not meet the stronger requirements.
These disclosures do not prove Astra is GPT-6 or that Bel is its base. They do show that OpenAI has a significant unreleased system in active evaluation. Astra’s eventual system card, benchmarks, pricing, and availability will reveal far more about OpenAI’s progress than Bel’s rumored parameter count.
Stargate’s Real Advantage Is Training Optionality
OpenAI and SoftBank originally presented Project Stargate as a plan to invest as much as $500 billion in US AI infrastructure. By April 2026, OpenAI said it had already surpassed its initial target of securing 10 gigawatts of capacity by 2029, including more than 3 gigawatts added during the preceding 90 days. The company also confirmed that GPT-5.5 was trained at Stargate’s Abilene facility using Oracle Cloud Infrastructure and Nvidia GB200 systems.
The immediate benefit is more capacity for ChatGPT and API inference. The deeper strategic benefit is optionality during model development.
A lab with excess compute can run more architecture experiments in parallel, test different data mixtures, recover from failed training runs, and allocate more resources to reinforcement learning, synthetic-data generation, evaluations, and safety testing. It can also reserve one cluster for a frontier pretrain without taking as much capacity away from existing products.
This changes the economics of research. A rival may have the ideas and talent required to build a competitive LLM but still have to delay the project while it waits for chips, networking equipment, power, or a sufficiently large synchronized cluster. Compute capacity affects not just how big a model can become, but how quickly a laboratory can learn which design is worth scaling.
Stargate’s headline capacity should not be confused with one operational supercomputer. Planned or contracted gigawatts may be split among regions, hardware generations, cloud providers, training workloads, and inference services. Even so, OpenAI has moved beyond plans: Abilene has already trained at least one frontier model, and the company says additional Stargate sites are supporting next-generation research.
Anthropic’s Compute Gap Is Plausible but Unproven
Saying Anthropic “lacks compute” would be misleading in absolute terms. The company has announced some of the largest AI infrastructure agreements in the industry.
In May, Anthropic said a SpaceX partnership would give it access to more than 300 megawatts of capacity and over 220,000 Nvidia GPUs at the Colossus 1 data center. It has also secured up to five gigawatts from Amazon, including nearly one gigawatt expected by the end of 2026, and five gigawatts of Google and Broadcom capacity scheduled to begin coming online in 2027.
Those figures do not necessarily refute the Bel report. Much of Anthropic’s immediately available SpaceX capacity was presented as a way to increase Claude usage limits, suggesting that serving customer demand is consuming a meaningful portion of the new hardware. Meanwhile, a large share of its next-generation training capacity will arrive later in 2026 or during 2027.
The timing is notable. The report claims Anthropic expects OpenAI to hold the advantage through much of the rest of 2026 but believes it can regain the lead early next year. Anthropic’s publicly announced capacity schedule supports that possibility, although it does not confirm any internal model roadmap.
Anthropic is not standing still on models either. It released Claude Opus 5 on July 24 after launching Sonnet 5 in June. The absence of a publicly announced Astra competitor does not establish that no such model exists. It could indicate a smaller pretrain, a later training slot, an unfinished post-training program, or simply a different release strategy.
A more precise interpretation is that Anthropic may face a near-term frontier-training constraint, not a general compute shortage. The difference matters. Capacity used to keep Claude reliable for millions of users cannot always be reassigned to a months-long experimental run, and newly contracted gigawatts do not help until the hardware is installed, networked, and operational.
What Would Confirm the Bel Story
Several disclosures would move Bel from rumor to defensible reporting:
An OpenAI announcement or system card naming Bel or explaining its relationship to Astra.
Architecture details separating total parameters from active parameters.
Information about training tokens, compute, data composition, and post-training.
Reproducible evaluations comparing the resulting model with GPT-5.6 and Claude.
A clear explanation of whether Astra, GPT-6, and Bel are models, product families, or internal training projects.
Until then, “AGI-threshold” should be treated as roadmap language rather than a measured technical category. The report does not attach that phrase to a published evaluation, capability threshold, or deployment standard.
Final Thoughts
The strongest version of this story is not that OpenAI has already trained GPT-6 or quietly crossed an AGI threshold. The evidence does not support either conclusion.
The more defensible argument is that OpenAI may be entering a period in which infrastructure lets it attempt frontier training runs sooner, more frequently, and with fewer compromises than its rivals. Anthropic’s announced expansion could narrow that gap in 2027, but hardware arriving next year cannot train a model needed for a launch this year.
Astra will be the first meaningful test. If it converts OpenAI’s compute advantage into a clear improvement on long-running agentic work, science, coding, and reliability, Stargate will have produced something more valuable than abundant inference: a faster research cycle for each future generation of AI.
Frequently Asked Questions
5 questions
1
What Is OpenAI Bel?
Bel is the reported codename for a newly completed OpenAI pretraining run with more than 10 trillion total parameters. The report connects it to Astra, GPT-6, and later models, but OpenAI has not publicly confirmed the name, architecture, parameter count, training completion, or intended product. Bel remains an unverified internal-model rumor.
2
Is Bel the Same Model as GPT-6?
Bel is not confirmed to be GPT-6. The report suggests OpenAI could use Bel as a base for Astra or GPT-6 after reinforcement learning and additional post-training. OpenAI has acknowledged Astra as an upcoming model but has not said whether it will carry the GPT-6 name or use Bel as its foundation.
3
Does Bel Really Have More Than 10 Trillion Parameters?
The 10-trillion-parameter figure has not been independently verified. Even if accurate, it may describe total rather than active parameters, especially if Bel uses a sparse mixture-of-experts architecture. Without architecture details, training-token counts, compute information, and evaluations, the figure cannot establish how capable or expensive the resulting LLM will be.
4
Why Does Project Stargate Matter for GPT Models?
Project Stargate gives OpenAI more capacity to train models, run experiments in parallel, conduct reinforcement learning, perform safety evaluations, and serve existing products. Its advantage is therefore broader than faster ChatGPT inference. Large, available training clusters can shorten OpenAI’s research cycle and let it pursue frontier models without diverting as much infrastructure from customers.
5
Is Anthropic Running Out of AI Compute?
Anthropic is not running out of compute, but it may face a near-term shortage of capacity suitable for its largest training runs. The company has secured hundreds of thousands of GPUs and multiple gigawatts through SpaceX, Amazon, Google, Broadcom, and other partners. However, substantial portions of that capacity are supporting inference or will not become operational until late 2026 and 2027.
