ChatGPT's new Bel model has been revealed to be ready for training!

  It's the first time in the history of the world.

The tech world is completely sleepless!

OpenAI's most powerful model, the Astra, is expected to be released next Thursday and has expanded its test scope.

Previously, internal leaks have officially taken over the underlying code of the Jalapeño chip. AI writes the core code running 1.8 times faster than top human engineers.

OpenAI also has a deep-water bomb.

It is rumored that OpenAI has successfully run a 10 trillion parameter pre-training model with the code "Bel", pointing directly to the AGI terminal threshold, and even pre-training with more than 10 trillion parameters is just the beginning for Bel. After that, Bel can learn at two speeds.

Ottman has said: "We should throw another party for the next generation of models".

What is the ambition of the post-GPT-6 world?

This year, OpenAI appears to be trapped in a bottleneck of scale expansion and even an emergency "Red Code" alarm.

For the sake of centralized computing, OpenAI even cut out products such as Sora and AI browser Altas.

The next-generation AI model, Astra, has taken a number of mathematical breakthroughs and surprised the world, but it's never been released.

In recent weeks, there has been unrest in OpenAI's workforce: senior officials such as Chief Revenue Officer Dennis Dresher, Chief Operating Officer Brad Lightcap, and former Deputy Chief Executive Officer Sam Ottman, Phineas Seymour, have stepped down.

While the outside world has been speculating whether OpenAI is "out of the blue", a number of hardcore technological breakthroughs have erupted.OpenAI has just completed super-massive pre-training, code-named "Bel".

How scary is this "Bel"?

Let's look at a few keywords:

1. Break through the 10T (10 trillion) parameter key

GPT-4 with 1 trillion arguments gives AI the common sense and logic of almost human undergraduates, so how will Bel with 10 trillion arguments "emerge" in complex reasoning, long-text correlation, interdisciplinary multi-modal understanding?

It's a change in mass.

It's the equivalent of packing the brain capacity of all the world's top experts together and multiplying it by an exponential amplifier. Under this exponential level, AI's understanding of world models will reach an unprecedented depth.

2. The successor to "Doug", the ultimate baseline after GPT-6.

The outbreak pointed out that before that, OpenAI had already completed the pre-training of the code name "Doug".

DougIt is classified as an Astra project and the base model of the rumored GPT-6 (following it will also be matched by extremely high-intensity reinforcement learning).

As the successor, BelIt's a lot more than Doug."After the GPT-6 era".The next generation's foundation.

Bel is directly targeting the tech world's holy grail.AGI (General Artificial Intelligence)。

Bel models are reported to outperform Astra in coding, reasoning, and long-range intelligence tasks.

According to sources, the model can operate efficiently for several days without human intervention, recover itself and coordinate hundreds of parallel sub-intelligent subagents.

The Fable Killer (played by Claude Fable)

@ChrisGPT is frank in his tweet:

Bel is a "monster" class model of OpenAI designed to be a Fable Killer.

It is expected to be released before the end of this year, or within a few months of the release of Astra.

He even received the code six days ago, confirming the reliability of the source.

In theory, Bel may have exceeded GPT-6, or even approached the OpenAI-defined AGI threshold.

In the official release of GPT-5.6, they introduced the "RSI Index", which integrates results in research modeling, core and training formula optimization, machine learning experiments, and model self-improvement.

Sol subsequently designed hundreds of architectural experiments for its smaller draft models, launching training, and only intervening artificially in the event of hardware failures and unstable training.

Bel has become a truly continuously evolving entity.

The Fast Weight Layer will absorb experience lessons from verified proofs, code testing, experiments and tool trajectories during its work; while the slower cycle will consolidate the improvements that survive in the evaluation into perpetuation weights and training formulations.

It learns rapidly in fast memory, converts verifiable improved hardness to slow gravity, while continually optimizing the operating mechanisms for the next round of learning and distilling the final results into small models that can actually be used by everyone.

The GPT-7 may just be a security snapshot of Bel in a week's state.

On Reddit, Leo's news has sparked a lot of discussion, and after all, Leo's outbursts are still credible.

It has been speculated that the internal model may be 4.5-6 months ahead of the external model.

If Bel is a technological reduction hit, then the calculation is the ultimate trick of morale within OpenAI.

According to the cross-analysis of multiple trackers, OpenAI judges that the next half of 2026 to 2027 will see a significant increase in the number of tracking users.In the meantime, we're going to have to stay ahead of the curve.

Why are you so determined?Calculation。

OpenAI's internal assessment suggests that its biggest enemy, Anthropic, is a lack of computing power that makes it difficult to cope with the next generation of OpenAI.。

In the arms race of large models, arithmetic is a bullet. When the model parameters rise to 10 trillion levels, training at a time is costly.

While Anthropic is highly sophisticated in model architecture and alignment technology, it has obviously struggled in the face of absolute "violent aesthetics".

In the face of the upcoming public debut of Astra, Anthropic is limited to the mathematical bottleneck, and it will be difficult to give a strong response this year.

While other companies are still trying to figure out how to match 100,000 H100/B200 chips, OpenAI has come up with "Doug" and "Bel" with violent aesthetics.

With the release of OpenAI's self-developed Jalapeño chip, the Forbidden City River has not only not been flattened, but has been stretched.

10 trillion arguments may seem a bit distant. But within OpenAI, these underlying violent breakthroughs are ultimately pointing to the same AI endpoint.

Recently, in an interview with renowned tech podcast Matthew Berman, OpenAI executive Tibo unreservedly played through OpenAI's future roadmap at the application level.

He even exclaimed:

The Codex model, which is extremely powerful at the moment, will look like a product of a primitive era in the next two to three months.

In conjunction with the previous outbursts, Tibo points to four disruptive transformations.

Retro-self improvement (RSI) happens every day within OpenAI.

Tibo confirmed that the "internal weirdness" not only exists, but has already gone out of business.

OpenAI has long used the most powerful models to optimize its own reasoning, CUDA core and even infrastructure.

He gave an example: using the Sol model internally to optimize the Luna model, directly cutting operating costs.80%！

"Ultra Fast" will reshape the human heartbeat.

Tibo reveals that the current Ultra Fast mode in-house has been achieved14 timesHe predicts that in just one to two years, this ultra-low delay will become the industry standard.

When AI's response speeds approach or exceed human thinking speeds, interaction will become real-time. Workflow will completely return to the "flow state" and human cognitive load will decrease exponentially.

Calculation paradigm shift: Your computer is about to become "waste metal".

With the arrival of the next generation of models (such as Astra), Tibo makes it clear that the mainstream AI of the future is definitely no longer running locally, but rather that it's going to be in the future.Moving to the Cloud for a Large Group of Agents。

In the end, ChatGPT merged with the Codex to become Personal AGI.

In the future,"The mechanism will disappear completely".You don't need to write complex Prompt manually, you don't need to maintain skill files, you don't need to manage memory, you don't need to manually schedule Agent.

Personal AGI will continue to understand your goals, daily habits, and team dynamics passively.ActivelyIt's a great way to help.

What's even more amazing is that it'sDynamic UI AdaptationThe underlying clarity is the same set of multi-mode, Voice-first technologies, but the interface changes automatically like water depending on who you are.

Whether the rumors of "Bel" are exaggerated or whether Astra's hands tear the human expert code so smoothly, this massive outbreak has sent a signal to the world that should not be ignored:

The development of AI is not stalled, it is just saving up a storm enough to turn the tables.

In the second half of 2026, the good drama has just begun.