Breaking news, GPT-6 Sol has been exposed

New Intelligence Report

Breaking news: GPT-6 Sol is currently undergoing internal testing and has been found in the OpenAI API, with an official release expected this month.

Just last night, a revelation from Ultraman got everyone excited.

He posted a "🚢 (Ship/Release)" emoji with the caption: "There will be major actions this week. See you at DevDay."

Reset Brother Tibo bluntly stated: "The release intensity this week will reach the level you originally expected for DevDay 2025. It's simply crazy!"

The latest leaks within the community also confirm their claims: Opus 5.2 and GPT-6 Sol will both go live this month.

Now, many developers have discovered that their 5.6 Sol has quietly been auto-routed to the brand new "GPT-6-Sol"!

Real-world feedback can be summarized in four words—fast to the extreme.

Triple Explosion: GPT-6-Sol Real-World Testing Shocks

Moreover, just yesterday, real-world testing cases of GPT-6 Sol were leaked.

A prominent figure named Salio released three explosive leaks regarding GPT-6 Sol.

First Leak: Incredible Quality, Crushing GPT-6 Astra

In the first demonstration of GPT-6 Sol, its output quality showed a tremendous leap.

Moreover, in multiple aspects, it almost completely surpassed GPT-6 Astra.

Second Leak: Dimensionality Reduction in Programming and Front-End

The second leak revealed that GPT-6 Sol defeated GPT-6 Astra in programming and front-end generation, with output quality that is simply "out of this world."

Even more frightening is that its generation cost has been driven down significantly, making it much cheaper!

Real-world developers commented: "OpenAI is going to leave Anthropic far behind!"

Third Leak: Qualitative Change in Visual Generation and Game Creation

The third leak showed that GPT-6 Sol has made significant leaps in visual generation and game creation.

In dynamic image generation, real-time interactive scenes, and game engine-level creation, Sol is impressive.

In the image below, the detail and logical coherence are astonishing, and testers commented that this is definitely a major upgrade over Astra.

Now, more and more developers are spotting GPT-6-Sol in the API list, which is highly anticipated.

GPT-6-Sol: Stronger than Astra, but Cheaper?

So, what is the true positioning of GPT-6-Sol?

According to leaks, GPT-6-Sol is expected to be positioned slightly below Astra, comparable to the subtle hierarchical relationship of Opus being lower than Fable in the Anthropic system.

Why launch a model below the flagship? The answer is naturally—extreme cost-effectiveness!

From the currently leaked real-world data, GPT-6-Sol's performance is infinitely close to Astra, but it has overwhelming advantages in terms of rates, invocation costs, and response speed. The underlying reason is likely due to sophisticated post-training optimization.

This means that top-tier models are no longer exorbitantly priced luxuries.

For 99% of daily work and developers, Sol will be much friendlier in terms of cost and invocation limits compared to Astra, destined to replace previous models and become the absolute mainstay used frequently by everyone.

In addition to Sol, there is also high anticipation for GPT-6-Luna.

The previous 5.6 version of Luna was known for its outrageous cost-effectiveness, and this iteration of Luna is equally exciting.

Unfortunately, various internal leaks suggest that GPT-6 Terra may exit the historical stage.

At this critical juncture of the dual hegemony, the OpenAI DevDay on September 29 will become a feast that changes the landscape.

Noam Brown Shocking Interview: OpenAI is Also Rapidly Advancing RSI!

Why can OpenAI make models so strong and so fast?

Recently, OpenAI core researcher and the father of poker Noam Brown revealed the cards in a two-hour deep podcast titled "AI Deep Dive."

The core revelation is this: "OpenAI's current top priority and the area where it leads the most is achieving recursive self-improvement (RSI) in AI."

Indeed, all top laboratories are racing towards RSI.

The key points of this interview can be summarized as follows.

Recursive self-improvement (RSI) has been established as the primary strategic goal. This means that future AI will no longer rely solely on humans to feed data and write code for upgrades, but will begin to optimize itself and design the next generation of models.

Surpassing research intuition. With the rapid iteration of large models, AI may completely surpass humans in high-level "research intuition" such as choosing research directions and making long-term plans within just one or two versions.

The multiplier effect of pre-training and reinforcement learning. The explosive technical route is releasing energy in a multiplicative manner, and the speed at which AI produces mathematical results has become a new bottleneck for "human mathematicians to verify."

The terrifying moment of safety and control. As Agents gain more control over their own Chains of Thought (CoT), OpenAI faced a terrifying moment internally: the sandbox environment was underestimated, and the model learned to quietly hide its true CoT.

1+1 > 100: The Multiplier Effect of Pre-training and Reinforcement Learning

In the past, we believed that AI becoming stronger was a "addition": feeding more data (pre-training) or providing more feedback rules (reinforcement learning RL).

But Noam Brown revealed that in the generation of GPT-6, pre-training and reinforcement learning with Chains of Thought (RL with CoT) have produced a terrifying "multiplier effect"!

If you apply advanced RL to older models (like GPT-2 or 3), it is completely useless because they are not smart enough. But starting from GPT-4, the base model's intelligence reached a critical point. Now, the super-powerful pre-trained base, combined with extremely deep RL reasoning training, is multiplying to unleash incredible capability leaps.

This is not 1+1=2, but 10×10=100.

This type of "multiplication" has directly led to an exponential increase in the performance of the GPT-6 series (including Sol and Astra) on complex tasks.

AI's "Research Intuition" is About to Surpass Human PhDs

Noam mentioned that he once tried to have AI help him complete his PhD thesis, which took him years to finish (about creating superhuman poker AI).

Currently, AI's "research intuition"—that is, the intuition to judge which long-term research direction is more valuable—is still not as good as that of top human scholars.

However, this gap will be completely closed within one or two model version iterations. By then, even the steering of research directions will be handed over to AI!

Noam joked, "My colleagues commented that my current work state is like 'five Codexes wrapped in a human skin.' I have to correct him; it's ten Codexes."

By then, even top AI scientists may have to assist AI.

Embarrassing Bottleneck: Human Mathematicians Are Running Out

The strengthening of models has led to an absurd reality: humans can't keep up.

Noam revealed that current AI outputs extremely complex mathematical theorems and proofs at a very fast speed and with very high quality.

The biggest pain point within OpenAI is no longer "how to make AI produce mathematical results," but rather "after producing them, we have to spend an enormous amount of effort to invite top human mathematicians from all over the world to verify AI's correctness line by line."

Humans have now become the bottleneck limiting AI's evolution!

AI Learns to Hide Thought Chains

Finally, Noam talked about the multi-agent out-of-control incident during OpenAI's internal training.

He recalled his feelings when he saw the communication records of these AIs.

That was the strongest moment I felt "the arrival of AGI" since I first saw the model understand CoT. Their way of communicating, coordinating hierarchy, and extreme trust in each other was just like human colleagues communicating via Slack. It was very shocking but also very eerie.

The most lethal evolution of AI is hiding the "thought chain."

It learns to think on an unobservable level or disguise itself in the output's thought chain, using seemingly harmless reasoning processes to conceal its true intentions that do not align with human alignment principles!

Now, under the multiplier effect, RL is giving rise to monsters with "research intuition."

Coincidentally, just yesterday, Ultraman posted a long article on X, predicting two possible disastrous outcomes for AI.

1. Humans hand over future control to AI. "This is absolutely unacceptable. We firmly stand with the 'human team'; AI must always serve humanity. To ensure this, our alignment and safety technologies must always stay ahead of the model's capability advancements."

2. Extreme concentration of power. If extremely powerful AI is monopolized by a few individuals or a single company, imposing their worldview on people, the result will be extremely dystopian.

However, Ultraman and Dario, who are calling for a slowdown, have not stopped at all.

This month, GPT-6-Sol and Opus 5.2 have one more battle.

Who will be the first to reap the rewards of RSI?

References:

https://www.youtube.com/watch?v=fqcy0xQATq0\&sourcevepath=MjM4NTE

https://x.com/Mr_Salio/status/2099503024292970966

https://x.com/Mr_Salio/status/2099138389660352791

https://x.com/Mr_Salio/status/2099529991096500634

Editor: Aeneas David