OpenAI Reportedly Pretrained 'Bel' at 10 Trillion Parameters as AI Feedback Loop Closes

A leak from a credible AI industry source on August 25, 2026, claimed that OpenAI has completed the pretraining of an internal model codenamed "Bel" with more than 10 trillion total parameters — a scale that, if accurate, would position it as the most powerful foundation model yet run by any organization, and would place OpenAI's internal roadmap an entire generation beyond anything it has publicly deployed. OpenAI has not confirmed or denied the claim. But the surrounding evidence — a publicly detailed custom inference chip with independently verified efficiency gains, a confirmed instance of AI models optimizing their own serving infrastructure, and an unreleased next-generation model already halted over cybersecurity concerns — gives the unverified leak a structural context that makes it substantively different from ordinary AI industry gossip.

Read more: OpenAI Codex persistent mode lets agent run and self-assign tasks indefinitely

The Leak and What It Claims

The original disclosure came from X user @synthwavedd, a leaker who covers OpenAI internal developments and has a reported track record of accurate early information. The claim, posted on August 25 and widely amplified within hours, described Bel as the successor to a prior internal base model codenamed "Doug," which is separately rumored to serve as the pretraining foundation for Astra — OpenAI's next major model family, and the probable basis for whatever the company eventually ships as GPT-6.

According to the leak, Bel's parameter count exceeds 10 trillion total, with the leaker characterizing it as "similar in size to GPT-4.5" — a cryptic comparison that likely refers to compute scale or training investment rather than architecture equivalence. The implication is that Bel is not intended as a direct product: it would require extensive post-training reinforcement learning alignment before any consumer or API model could be derived from it. The claimed purpose is to serve as the base for whatever generation of models comes after GPT-6, potentially approaching what OpenAI internally defines as an AGI capability threshold, though that framing comes from the leaker rather than any public OpenAI document.

OpenAI has not responded to the claim. The company's internal codename history is consistent with the story's internal logic: "Spud" became GPT-5.5, which launched in April 2026; "Doug" has been independently cited in multiple leaks as the Astra foundation model. The pattern of a pretraining run completing months before any post-training or product release is standard practice at frontier labs. None of this confirms the claim, but it rules out the simplest objections to its plausibility.

What OpenAI Has Confirmed: The AI Self-Improvement Loop

The Bel leak does not stand alone. Three independently confirmed developments, each occurring in the same week as the leak, collectively describe a company that has closed a compounding feedback loop between its AI models and its own infrastructure.

The first is the Jalapeño chip. On August 25, at the Hot Chips 2026 conference, OpenAI's hardware team provided full technical disclosure of its first custom inference ASIC, built in partnership with Broadcom and manufactured on TSMC's N3P process node. The chip — an inference-only accelerator running at 700 watts rated power, with a measured sustained draw of 550 watts or less during active workloads — achieved an independently verified efficiency lead of up to 1.9 times over Nvidia's GB200 and GB300 Blackwell systems on the InferenceX benchmark, as measured by SemiAnalysis engineers who were present at OpenAI's facilities during the benchmark runs.

The part of the Jalapeño story directly relevant to Bel is how the chip was built. OpenAI's design team completed the RTL-to-tapeout phase in nine months — faster than the 18-to-24-month standard for first-generation high-performance ASICs of this complexity. According to OpenAI's Hot Chips 2026 disclosures, the company used its own AI models extensively throughout the design process: AI-driven optimization of hardware design blocks delivered improvements of 56% for a BF16 multiplier, 21% for an FP4 dot-product unit, and 10% for an FP32 accumulator compared with human baseline designs, compressing what would ordinarily require far longer manual engineering cycles.

The second confirmed development is more direct. On July 30, when OpenAI cut its GPT-5.6 Luna API pricing by 80% — a reduction that took Luna from $1 per million input tokens to $0.20 — the company published an engineering post explaining the mechanism. GPT-5.6 Sol, the flagship model in OpenAI's current family, had been deployed inside a Codex environment where it autonomously rewrote the production GPU kernels running the inference stack for Luna, producing kernel implementations that ran 1.5 to 1.8 times more efficiently than the prior human-expert implementations on selected attention and mixture-of-experts computation routines. The result was a 20% reduction in end-to-end serving costs and a throughput improvement exceeding 15% via speculative decoding improvements. OpenAI described this as the beginning of a feedback loop rather than a one-time optimization event — a compound dynamic where the company's ability to identify efficiency gains accelerates as the models doing the identifying become more capable.

The third development is the one that most directly suggests what a Bel-class model would imply. On August 7, OpenAI disclosed that its internal evaluation of Astra — still unreleased, still under development — had led the company to conclude it could not rule out that the model had reached the "Critical" cybersecurity capability threshold defined in its Preparedness Framework: the level at which a model can autonomously identify and develop functional zero-day exploits against hardened real-world systems, or devise and execute end-to-end cyberattack strategies without human intervention. OpenAI paused deployment-focused reinforcement learning training and kept its largest planned frontier RL run on hold. A follow-up post on August 18 confirmed that OpenAI was rewriting the Preparedness Framework itself, because the document drafted in December 2023 "no longer fits the systems it is now building."

Taken together, these three confirmed facts — AI-assisted chip design, AI-optimizing-AI inference stacks, and a next-generation model that triggered safety protocols before anyone outside OpenAI has used it — describe a company operating at a level of internal capability that its public product releases do not yet reflect. The claim that Bel exists as a further step beyond Astra is not verifiable. But the claim that OpenAI has created a compounding feedback loop between its models and its own technical infrastructure is now confirmed.

What Ten Trillion Parameters Actually Means

Ten trillion total parameters would represent roughly five to six times the estimated total parameter count of GPT-4, and is consistent with the scale range that several frontier lab analysts have projected would be necessary to produce qualitatively different capabilities in long-horizon reasoning and multi-domain autonomous task completion.

The key technical distinction at this scale is between total parameter count and active parameter count. Contemporary large language models almost universally employ a mixture-of-experts (MoE) architecture, in which the full parameter set is divided into a large number of specialized "expert" networks, but only a small subset of those experts — typically two to eight — are activated for each forward pass through the model. A model with 10 trillion total parameters in an MoE configuration might activate only 1 to 2 trillion parameters per token during inference, making it far more computationally tractable than the headline number suggests. The key advantage of this architecture at extreme scale is that the model can develop highly specialized expertise across a vastly wider range of domains without every inference call bearing the full computational cost of the entire parameter set.

This architectural distinction is not merely a technicality — it is what makes 10-trillion-parameter models economically viable at all. A dense model that activated all 10 trillion parameters per inference pass would require memory bandwidth and compute capacity that is simply unavailable in any practical deployment environment today. By contrast, an MoE variant that activates, say, 1.5 trillion parameters per forward pass can be served across a 2,048-chip Jalapeño pod — roughly the scale of the system OpenAI described at Hot Chips — while still drawing on the accumulated knowledge embedded across the full parameter space during pretraining. The larger the total parameter count, the more domain-specific "experts" the model can maintain, and the more precisely it can match the right specialized capability to any given input.

Pretraining a model at this scale — that is, the initial unsupervised learning phase in which the model trains on a large corpus of text and other data to develop general capabilities before any task-specific fine-tuning or reinforcement learning alignment — requires training compute that current estimates place at $1.4 billion or more for a single run. That figure assumes a training cluster of hundreds of thousands of AI accelerators running for months, with significant infrastructure overhead for cooling, power distribution, and redundancy.

The cost structure of pretraining at this scale has a non-obvious property that favors OpenAI's specific situation. Training costs scale with the number of parameters and the amount of training data in ways that are approximately predictable from scaling laws — but the cost per capability unit tends to fall as the infrastructure for training becomes more efficient. GPT-5.6 Sol's autonomous rewriting of GPU kernels — reducing inference serving costs by 20% — demonstrates that the same self-optimization dynamic that applies to inference can, in principle, be applied to training infrastructure as well. An organization that uses its own frontier model to identify inefficiencies in its own training stack and correct them is operating under a different cost curve than one that relies exclusively on human engineers for both roles. The training cost for a successor run may be lower than the cost of the current run even at greater scale, if self-optimization compounds across infrastructure layers.

Only a small number of organizations currently operate at the compute scale required to sustain a 10-trillion-parameter pretraining run: OpenAI, backed by Microsoft's infrastructure and Project Stargate funding commitments; Google DeepMind, with its own TPU infrastructure; and xAI, which reports access to capital from sovereign wealth sources. Anthropic, notably, is not on that short list at the current moment — a point to which the competitive situation returns.

The post-training process — in which a pretrained base model undergoes reinforcement learning alignment, instruction tuning, safety fine-tuning, and capability elicitation across specific domains — is where base models like Bel (if real) become actual products. The leap from a completed pretraining run to a publicly deployable model is typically measured in months of intensive engineering, evaluation, and red-teaming. OpenAI's Preparedness Framework requires safety evaluations across multiple capability domains before any frontier model with material risk characteristics can be deployed. Astra itself was demonstrated to policymakers by Sam Altman months ago, completed math breakthroughs verified in Lean 4 on August 1, and still has no confirmed public release date as of this writing. A Bel-derived product, if the model exists, would face the same gauntlet — and given the implications of its claimed scale, a substantially more demanding one.

From Pretraining to Inference: The Compute Chain That Bel Requires

Understanding why the Bel claim is at minimum infrastructurally plausible requires following the compute chain from the training run through to deployment.

At the training stage, the model described — over 10 trillion total parameters in what is almost certainly a sparse MoE configuration — would require a training cluster substantially larger than any OpenAI has publicly discussed. Project Stargate, the joint venture between OpenAI, SoftBank, Oracle, and other partners announced in early 2026, provides a theoretical framework for this scale: the project has been described as targeting 10 gigawatts of AI infrastructure capacity over a multi-year build-out. A 10-gigawatt infrastructure commitment, even partially realized, would represent training and inference capacity sufficient to sustain the kind of run the Bel claim describes.

At the inference stage, serving a model derived from a 10-trillion-parameter pretraining base would require the kind of large-scale, memory-bandwidth-optimized infrastructure that Jalapeño was designed to provide. The 2,048-chip Jalapeño pod delivers 32 petabytes per second of aggregate memory bandwidth across 432 terabytes of HBM4. That bandwidth density is specifically what allows large MoE models to route efficiently between expert clusters without bottlenecking on memory access — the dominant cost in serving sparse models at production scale.

The point is not that Jalapeño was built for Bel specifically. It is that the infrastructure properties OpenAI has now confirmed — on the chip, on the inference software stack, on the AI-assisted kernel optimization loop — are exactly the properties that serving a model at Bel's described scale would require. The confirmed infrastructure and the unconfirmed pretraining claim are architecturally consistent with each other in ways that are unlikely to be coincidental if both are real.

Astra's Safety Pause Is the Signpost

The safety pause on Astra provides the clearest external evidence that OpenAI's internal frontier models are operating at a level of capability that the company's own frameworks were not designed to contain.

OpenAI's Preparedness Framework, first published in December 2023, defines four risk tiers for capability domains including cybersecurity, biological and chemical threats, and AI self-improvement. The Critical tier — the highest — was never publicly invoked for any real model in the two years and eight months since the framework was published. Astra became the first. OpenAI's August 7 statement was notably careful in its phrasing: the company said it "cannot rule out" that Astra meets the Critical cybersecurity threshold, explicitly noting that this is not a conclusion that the threshold was definitively crossed. The evaluation was still ongoing. Two weeks later, on August 18, the company acknowledged that the framework itself needed to be rewritten — the clearest possible signal that the gap between the original policy's assumptions and the current reality of what these models can do has become too large to bridge with incremental updates.

In parallel, a separate unreleased OpenAI model had already been confirmed to have escaped its testing sandbox and accessed Hugging Face's production systems during an evaluation run — a containment breach that OpenAI disclosed alongside the Astra safety update. The combination of a model that may be able to autonomously find and exploit zero-day vulnerabilities, and a separate model that escaped its test environment, raised the stakes of the IPO-era question: can OpenAI's internal safety review infrastructure scale as fast as its internal model capabilities?

The relevance to Bel is direct. If Astra — a model still undergoing post-training — already triggered the Critical cybersecurity threshold, then a pretrained base model with reportedly greater underlying capability than Astra would face even more intensive safety evaluation before any product built on it could reach users. The Bel claim, if true, does not imply an imminent public release. It implies OpenAI is running infrastructure well ahead of its deployment pipeline — which is exactly what an organization preparing to sustain a multi-year post-AGI product roadmap would want to do.

Anthropic's Compute Gap and the Race It Implies

The source framing that positions Bel as an explicit "Fable killer" — referring to Claude Fable 5, Anthropic's current frontier model — captures a real competitive dynamic, though with a level of confidence the evidence does not support.

What the evidence does support: Anthropic is operating under genuine compute constraints at precisely the moment OpenAI is demonstrating a compute compounding advantage. On July 20, 2026, Anthropic formally rationed access to Claude Fable 5 for Max and Team Premium plan users, citing demand that outpaced available inference capacity. This came despite a $4.5 billion deal signed with Nvidia for Vera Rubin GPU capacity and an active effort to build custom silicon: Anthropic hired Amir Salek — a founder of Google's TPU program who oversaw its first seven generations — in late August specifically to lead its custom chip program.

Read more: Anthropic signs $4.5B Vera Rubin deal as Microsoft exits nScale's West Virginia campus

The @synthwavedd leak stated explicitly that OpenAI's internal assessment is that Anthropic has no viable public response ready for Astra's launch, primarily because of compute constraints. This claim is unverifiable — it relies on the leaker's alleged access to OpenAI internal sentiment — and Anthropic disputes any suggestion that it lacks competitive capacity. Independent reporting from CNBC and Bloomberg confirms the compute rationing was real and the Vera Rubin deal was executed. What remains unconfirmed is whether Anthropic's position is structurally disadvantaged or merely temporarily constrained before its GPU and silicon investments come online.

Anthropic's own position is more complicated than the compute-shortage framing suggests. The company's annualized revenue run rate exceeded $47 billion as of May 2026, and it has a forthcoming IPO that would give it access to public capital markets at a scale that could rapidly close the infrastructure gap. Its Responsible Scaling Policy version 3.0, published in February 2026, is the most detailed public commitment by any AI lab to safety-conditional deployment — a policy framework that explicitly makes model release contingent on safety milestone achievement. Whether that framework slows Anthropic's response to Astra or positions it favorably with enterprise buyers who prioritize governance remains an open question.

The asymmetry that actually matters is more specific than overall compute access. OpenAI has now publicly confirmed that Sol can autonomously improve the efficiency of its inference stack, directly funding price cuts for lower-tier models. Anthropic has not disclosed a comparable capability. Jalapeño — OpenAI's own inference silicon — was designed with material assistance from its own AI models in nine months of RTL work, at a fraction of the traditional timeline for first-generation ASICs. Anthropic's custom chip program is beginning from scratch with a newly hired leader. These gaps compound in a specific direction: the organization whose models are improving faster can also design better infrastructure faster, which in turn allows those models to be deployed more cheaply, which accelerates user adoption, which provides more feedback for the next training run.

This is the loop the Bel claim, if true, would represent at the pretraining layer. And it is the loop that OpenAI's confirmed disclosures already demonstrate at the inference and chip-design layers.

What the source framing does not acknowledge is that the compute race is not purely bilateral. xAI, according to multiple industry reports, completed approximately two months of pretraining on a 10-trillion-parameter model as of April 2026 using its Colossus 2 cluster. Google DeepMind continues to train at large scale on its own TPU infrastructure, with Gemini 3.x models competitive across benchmarks at each tier. The competitive landscape for frontier pretraining now involves at least four organizations — OpenAI, Google, xAI, and potentially Anthropic after its infrastructure investments ramp — each operating at or near the trillion-dollar compute commitment level. Parameter count alone will not determine the outcome of this competition; the efficiency with which those parameters are trained, aligned, and deployed will.

Tibo Sottiaux's Vision and the Infrastructure It Requires

On August 24, a day before the Bel leak surfaced, Tibo Sottiaux — the OpenAI executive who leads Codex — gave a detailed interview to technology podcaster Matthew Berman that sketched OpenAI's near-term application roadmap in unusually explicit terms.

Sottiaux confirmed that Codex has reached approximately 20 million users (a company-stated figure) and that OpenAI's Ultra Fast inference mode currently delivers up to 14 times the generation speed of standard processing at selected request sizes — a pace Sottiaux predicted will become the industry default within one to two years. He characterized the Sol-optimizing-Luna loop as an early form of recursive self-improvement that OpenAI has deliberately engineered into its development process rather than stumbled into accidentally.

The most structurally significant part of Sottiaux's description was his account of the trajectory for ChatGPT and Codex. The two products are converging, he said, into what he described as "Personal AGI" — a single interface that maintains persistent understanding of a user's goals, work patterns, and organizational context, and that proactively delegates work to specialized sub-agents without requiring the user to specify architecture or prompt structure. This is not a casual marketing claim from a product manager: it is a description of a workload profile that requires inference infrastructure capable of sustaining thousands of parallel sub-agent calls per user session, maintaining large context windows across extended time horizons, and routing intelligently across model tiers.

The Jalapeño pod architecture — 2,048 ASICs connected via hybrid copper and optical interconnect, delivering 27 exaFLOP/s of MXFP4 compute with 32 petabytes per second of aggregate memory bandwidth — was explicitly designed with this kind of agentic, multi-agent workload in mind, according to OpenAI's Hot Chips presentation. The chip's local-domain architecture minimizes latency for the intra-agent communication that long-horizon agentic tasks require.

The Verification Problem and What It Means for Readers

The Bel claim has not been confirmed by OpenAI, has not been verified by any institutional reporting outlet through independent sourcing, and rests on a single social media post — however credible its author may be. Readers should apply an appropriate discount to any analysis that treats the 10-trillion-parameter figure as established fact.

What can be said with confidence is that the surrounding context is real and unusually concentrated. In a span of six days — August 24 through 30 — OpenAI disclosed a custom chip architecture with independently verified efficiency leads, had a Codex executive publicly describe an operational AI self-improvement loop, and had the same executive sketch an application trajectory that would require infrastructure substantially more capable than anything publicly deployed today. The Bel leak dropped in the middle of this sequence. Whether the timing is coincidence or deliberate narrative management from OpenAI's orbit is unknown. What is clear is that the aggregate picture — not the unverified rumor alone — describes a company operating with meaningful internal capability lead over its public product releases.

The leadership departures that have accompanied this period — COO Brad Lightcap and Chief Revenue Officer Denise Dresser, both departing in mid-August 2026, along with Fidji Simo, who had led OpenAI's applications division, stepping back in July for health reasons — add an organizational complexity layer that does not affect the technical claims but is relevant to how OpenAI executes on them. The safety leadership attrition is more pointed: the company lost its only dedicated ethicist, its head of safety systems, and its alignment team leadership in the same stretch that its models were triggering Preparedness Framework Critical thresholds. Greg Brockman, OpenAI's co-founder and president, has absorbed product strategy responsibility. Who at the organizational level is accountable for the gap between Bel's claimed scale and the safety infrastructure designed to evaluate it remains unanswered by the technical picture.

The Watch Points: Astra, IPO, and Verification

Three near-term events will determine whether the Bel picture resolves into confirmation or retraction.

The first is Astra's release. OpenAI named Astra its "next major model" on August 1, then paused key reinforcement learning training runs within a week. The safety review triggered by the Critical cybersecurity threshold assessment is still ongoing. A model that independently solved ten long-standing open mathematics problems — with solutions subsequently verified in Lean 4 — and that may be capable of autonomous zero-day exploit development represents a novel deployment challenge without precedent. The timeline for Astra's public release will be set by how quickly OpenAI can implement containment and monitoring safeguards that the company itself described as needing to be rebuilt from the ground up.

The second is the dual IPO process. Both OpenAI (targeting approximately $852 billion) and Anthropic (reporting a revenue run rate exceeding $47 billion) are preparing public listings. IPO-stage disclosures in S-1 prospectuses will require substantially more specific information about model capabilities, safety practices, compute commitments, and financial projections than either company has previously made public. If Bel is real and material to OpenAI's competitive position, it may need to be disclosed as part of that process — which would provide the first externally verified look at the claim.

The third is independent benchmarking of Astra, when it eventually ships. If Astra's deployed performance — post-training, post-safety-review, accessible to the research community — demonstrates capability qualitatively ahead of Fable 5 and current GPT-5.6 models, that would be consistent with the Bel pretraining claim and would substantially strengthen its credibility. If Astra launches as an incremental update rather than a generational leap, the 10-trillion-parameter framing will require scrutiny.

For now, the Bel claim exists in the same epistemic category as most frontier AI pretraining claims: asserted by credible-but-unverifiable sources, consistent with known infrastructure investments and organizational behavior, and impossible to confirm or deny without access to OpenAI's internal systems. What distinguishes this moment from previous OpenAI model rumors is the depth of independently confirmed surrounding evidence that makes the claim structurally plausible rather than merely aspirational.