OpenAI Allegedly Completes 100 Trillion Parameter Pre-training; Parameters Determine Capacity, Post-training Determines Price
Social media reports claim that OpenAI has recently completed a new round of pre-training internally codenamed "Bel," with total parameters exceeding 100 trillion; OpenAI has not confirmed the existence of Bel, the parameter scale, or the completion time of the training.
The reports describe Bel as the successor pre-training base to "Doug." According to this claim, Doug is expected to support Astra and GPT-6 level systems after further reinforcement learning, while Bel is positioned as the next-generation foundational model following that route.
The statement about "over 100 trillion parameters" does not disclose model architecture, the number of active parameters, expert mixture routing methods, training token scale, training set composition, training cluster scale, or training costs. Therefore, total parameters cannot be directly equated to the parameters actually used in a single inference call, nor can the model's capabilities be judged based on this.
Reports suggest that Bel may form a model after GPT-6 through reinforcement learning and could become the basis for the so-called "AGI threshold model"; the "AGI threshold" is not a product specification, model name, or verifiable performance metric publicly released by OpenAI, but rather an internally rumored positioning.
OpenAI has publicly confirmed that Astra is one of its "next major models," but has not announced a release date, price, model ID, or complete technical specifications. The company has stated that some planned cutting-edge reinforcement learning training is on hold due to cybersecurity risk assessments; this means that even if a pre-training has been completed, it cannot be inferred that related products will be released soon.
Regarding Anthropic's claim that "it is difficult to produce a model of the same level this year, mainly limited by computing power," this also originates from unverified social media reports. Currently, there is no official statement from Anthropic, disclosure of computing power procurement, or independent technical reports to prove that it cannot train or release a model of the same level as Astra.
In market mechanisms, such rumors themselves will raise expectations for cutting-edge training computing power, HBM memory, network interconnections, data center power, and model post-training data; the narrative benefits are concentrated on chip, cloud service, AI power, and data suppliers. The actual flow of funds still depends on whether OpenAI discloses training scale, deployment plans, and commercialization paths; competing companies bear the narrative pressure of "computing power lagging behind," but there is no verifiable evidence that the competitive landscape has been rewritten by Bel.
Source: Public Information
ABAB AI Insight
OpenAI's past product path indicates that "completion of pre-training" does not equal "completion of the model." After GPT-4, the company has separated model capabilities, reinforcement learning, tool invocation, multimodal interfaces, inference computation, and security assessments into different layers; the release of GPT-4o also shows that product naming does not have to correspond directly to the base training code. Therefore, even if there is an internal connection between Doug, Bel, and Astra, one cannot directly infer the final product performance or launch rhythm from an unverified training code.
In terms of capital pathways, if the 100 trillion-level "total parameters" is true, it most likely corresponds to an expert mixture architecture: the company covers different capability clusters with a vast total capacity, but each token is routed to only a portion of experts to reduce inference computation. The truly expensive aspects are not only the pre-training GPU duration but also high-bandwidth memory, cluster networking, data cleaning, reinforcement learning rollout, evaluation, and security red teaming. OpenAI's placement of the post-training phase of its highest capability Astra under a cybersecurity readiness framework indicates that capital investment must pass through deployment risks and commercialization thresholds to convert into revenue.
Historical comparisons can be seen in the "base-post-training-productization" competition among Google, Anthropic, and OpenAI: the original model scale provides the capability ceiling, reinforcement learning and inference computation determine the usability of complex tasks, and cloud distribution determines the speed of monetization. Even if Anthropic does not have a model of the same parameter scale as rumored, it may still compete through stronger post-training, coding agents, enterprise integration, and long-context capabilities; simplifying competition to parameters or a single training track overlooks model deployment quality and customer workflow barriers.
This belongs to capital concentration. Cutting-edge model training pushes the competitive threshold toward a few institutions that can long-term secure chips, build data centers, obtain power, and raise massive capital; parameter scale is merely a visible representation of this centralization. The mechanism is that the larger the training cluster, the higher the fixed costs of supply chain bargaining, data governance, security compliance, and post-training trial and error, making it difficult for small labs, even with research capabilities, to continuously pay the full costs from training to deployment.
ABAB News · Cognitive Laws
Parameters determine capacity, post-training determines price
The larger the model, the more competition resembles infrastructure rather than software
Unreleased capabilities cannot be counted as realized advantages