GPT-6 with 10 trillion parameters is set to be released this month!
Just now, an explosive message came from X.
AI investigative journalist ChrisGPT reveals that, despite government scrutiny, GPT-6 (Astra) will still be released in August.
GPT-6 is the model Astra, which OpenAI urgently paused yesterday.
Nine months ago, OpenAI began ramping up efforts, reallocating computing resources. Realizing that Anthropic was ahead in pre-training, OpenAI got worried!
GPT-4 is estimated to have approximately 1.8 trillion parameters. If Astra’s parameters are truly 10 trillion, that would be more than five times that of GPT-4.
Regardless, OpenAI has invested an unprecedented scale in pre-training.
Don't rush—there are still experts ahead.
The second leak from ChrisGPT is the year-end "epic beast" codenamed Doug.
It is not only OpenAI's largest pre-trained model to date, but also reportedly makes Fable seem primitive.
And it is not GPT-6.
However, this model requires pre-training followed by extensive cybersecurity testing, so it may be released no later than November.
Coincidentally, just recently, OpenAI President Greg shared a post on X—
GPT-4 was trained on August 8, four years ago.
Industry speculation suggests that GPT-4 had 1.8 trillion parameters; four years later, GPT-6 reached ten trillion parameters, achieving a terrifying leap in scale!
It can be inferred that GPT-6 will exhibit unimaginable logical reasoning and generalization capabilities.
Year-end ultimate beast Doug:
Compare Fable 5 to "primitive humans"
The name Doug was not only revealed by ChrisGPT but was also previously mentioned by the well-known firm SemiAnalysis.
According to leaked information, Doug's hashing power infrastructure is extremely premium.
Leading industry influencer Pankaj Kumar pointed out that Doug most likely trained on NVIDIA’s next-generation Vera Rubin chips, which have not yet been widely deployed!
Overall, OpenAI is playing a very large game.
In August, start by unleashing GPT-6 with 10 trillion parameters to make a splash and capture users' attention; by year-end, deploy Doug—the ultimate giant—to claim the undisputed number one spot.
Tibo, who is currently battling Boris, also teased a surprise.
In August, the decisive battle for ASI
In July 2026, multiple sources reported that OpenAI’s upcoming next-generation model would adopt a new pre-training foundation—surpassing the previous ~4T token “Spud”—with a scale of approximately 10T tokens and a context length potentially reaching 1.5M tokens, positioning it as a competitor to Anthropic’s Mythos/Fable.
The situation is clear now: Astra’s release has been delayed due to a security hold, but Anthropic has no intention of waiting.
Anthropic is preparing to release Fable 5.1 in August, priced the same as Fable 5, directly competing with Astra.
Anthropic is likely employing a "Tian Ji's horse racing" strategy: Opus 5 was deliberately held back to reserve Fable 5.1 as a surprise counter, ready to respond immediately upon GPT-6's launch.
Who will become the king of large models? This question became unprecedentedly intense in August.
Breaking Through the "Pretraining Wall": OpenAI's Counterattack
Why is Doug so important?
Because it points to a reality the entire industry dares not confront: since the release of GPT-4o in May 2024, OpenAI has not completed full-scale pretraining of its next-generation frontier model for over two years.
o1, o3, GPT-5 to GPT-5.5 are all essentially fine-tuned, reinforced, and inference-optimized versions built on the GPT-4o foundation. The "engine" of GPT hasn't been replaced in two years.
Due to severe pre-training bottlenecks, projects like Orion have been downgraded due to underperforming expectations.
But! OpenAI has never abandoned this core focus on pre-training.
Gemini 3, released at the end of last year, completely convinced Altman to switch his "engine."
Immediately afterward, a new pre-trained model codenamed "Garlic" emerged. According to Chief Research Officer Mark Chen, the company had resolved key issues in pre-training.
But Garlic was more of a proof-of-concept; Doug is the true product scaled up to a larger scale.
They likely overcame the performance degradation bottleneck and ensured the stability of clusters with tens of thousands or even hundreds of thousands of GPUs, leading to Doug's creation.
For a long time, SemiAnalysis has emphasized: Scaling Laws still hold.
o1 Series Proof: The model can continuously scale during inference time.
Meanwhile, Doug, who has made a triumphant return, declared: On the fundamental dimension of "pre-training," as long as computing power and data keep pace, Scaling Laws can still work miracles!
Only two of the big three remain
Now, only two of the Big Three remain.
On August 5, a part of Google's AI landscape collapsed.
After 27 years as Chief Scientist, Jeff Dean left with three core researchers to start a new venture. On the same day, Nobel laureate Hassabis stepped down as CEO of DeepMind and moved to the role of Chairman. Alphabet’s stock fell more than 4% that day.
DeepMind is no longer just a cutting-edge lab. The place that invented the Transformer is now supplying top researchers to its competitors.
Not long ago, Silicon Valley was home to the "Big Three" in AI, but now only OpenAI and Anthropic remain on the field.
Currently, Anthropic holds Fable 5. OpenAI holds the continuously iterated versions 5.5 and 5.6, as well as the upcoming version 6.
Currently, OpenAI holds two cards: Astra in August and Doug by year-end.
Over the past two years, OpenAI has shown that the old engine can still deliver much more power. When the engine is upgraded, how fast can this system, already pushed to its limits, take the car?
The answer to this question will be revealed by Doug at year-end.
Reference materials:
https://x.com/ChrisGPT/status/2086220662264250764
https://newsletter.semianalysis.com/p/gemini-is-cooked-but-gcp-is-cooking
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation, edited by Aeneas Solomon.