Just moments ago, an internal podcast by SemiAnalysis revealed several explosive pieces of news.
It is reported that Anthropic's new model, Mythos 2, has completed training but has been unceremoniously shelved.
Now, Anthropic is using data generated by Mythos 2 to go all-out on sprinting towards Mythos 3!
Anthropic's Most Powerful AI Neutered, Mythical AI Becomes Global Prohibited Item
In the podcast, well-known analysts Dylan Patel and Jordan Nanos jointly disclosed insider information—
The world's most powerful, top-tier AI models have long been fully trained, but they are being held down in the lab by the heavy hand of regulation, becoming internal prohibited items.
Many people believe that the moats of OpenAI and Anthropic have already collapsed.
SemiAnalysis tells us: This is all just an illusion.
It's not that open source is running too fast, but that the leading closed-source models are being forcibly shackled.
Jordan Nanos stated:
Clearly, we see this evident in the comparison between Mythos and Fable.
I simply cannot use Mythos, I can only use Fable, and sometimes I have to politely beg it to even get access.
As is well known, Anthropic internally still harbors the ultimate beast, Mythos, accessible only to enterprises participating in the "Glass Wing Project."
Mythos represents Anthropic's strongest technological "complete form." But it is too powerful, so powerful it induces panic, potentially even possessing disruptive cyber attack or autonomous evolution capabilities.
Therefore, Mythos is strictly restricted and absolutely not allowed to be released to the public!
Consequently, Anthropic had to fall back, wildly neutering and downgrading Mythos's capabilities, adding filters, before finally releasing a downgraded version—Fable.
The "advanced AI" we use daily is just a compromised product released by giants to appease regulators.
Asking for It and Getting It: Anthropic Reaps What It Sowed
Where did this trend of turning top-tier models into "internal prohibited items" come from?
This brings us to one of the most surreal scenes in Silicon Valley history.
During the livestream, Dylan Patel mercilessly mocked Anthropic:
I think the current situation is quite interesting, right? For years, the folks at Anthropic have been loudly calling out: Regulate us! Please, come regulate us! AI is too dangerous!
And now, suddenly, they've actually managed to scare everyone. Well, now Anthropic themselves can't release their models anymore. From what I know, Mythos 2 has completed all its training, but they just cannot release it.
It's arguably the biggest boomerang in AI history.
Led by Dario Amodei, the "AI safety fundamentalists" have long played the role of "whistleblowers." They've been roaming Capitol Hill, spreading apocalyptic warnings about AI destroying humanity everywhere, striving to push for stringent regulatory legislation.
Their original intent was perhaps to raise the regulatory bar, build a moat, and kick out small companies and open-source communities incapable of performing safety alignment.
In the end, they played with fire and got burned.
It is reported that Mythos 2, already trained, has demonstrated powerful code generation, logical reasoning, and cyber offensive/defensive capabilities, directly triggering the highest alarms.
Thus, Anthropic's Mythos 2, costing hundreds of millions of dollars in computing power, remains locked away internally.
Breaking the Feedback Loop: The Most Critical Crisis Has Already Arrived
Is not releasing a model really just as simple as making a bit less money or delaying its opening a bit?
Wrong! Dylan Patel pointed out the most fatal, most core crisis in this entire storm—the AI's self-iterative feedback loop might be cut off as a result.
"The really important question now is the internal feedback loop. Are they (by not releasing it) preventing themselves from using the internal Mythos 2 to make Mythos 3 stronger? Or, are they preventing themselves from using Astra to build the next-generation Astra Plus One?"
Previously, many technological leaps relied on this flywheel: Release a model ➡️ Collect massive amounts of real user feedback data ➡️ Use this high-quality data to train the next-generation model.
But now, Mythos 2 and Astra have been forcibly "locked down," and the flywheel grinds to a halt. AI companies lose the invaluable data from billions of real global interactions.
If Mythos 2 cannot be placed in the real world, what will Anthropic use to train the secretly-in-development third-generation beast, Mythos 3?
The answer is, apparently: They've initiated an internal "building gods behind closed doors" Ouroboros mode.
According to Patel, despite Mythos 2 being shelved, Anthropic's internal construction of Mythos 3 "has not stopped for a moment." The company's strategic focus has shifted from "rapidly pushing out new releases" to an extremely secretive "internal system improvement."
They are using Mythos 2 to generate data, write code, going all-out on the sprint towards Mythos 3.
Previously, Anthropic admitted in a risk report: The company internally runs a model stronger than Mythos 5, codenamed Model 2, and currently has no plans to release this model publicly.
On AEC: Model 2 scored 162.79. On CoBench, Model 2 scored 62.8%.
The external flywheel has stopped, but the internal flywheel is fully wound—computing power that once served hundreds of millions of users is being fully diverted into the laboratory.
Disconcertingly, three layers of recursion have already begun to interlock.
First Layer: Data Recursion.
Right or wrong is left for the world to decide—whether code runs or proofs hold does not require human annotation. Strong models set problems and solve them, verifiers grade, and data feeds directly into the next generation. This is self-play: AlphaGo Zero, with zero human game records, reached the top in 40 days. Anthropic has already admitted that Model 2 is generating large amounts of data.
Second Layer: Labor Recursion.
The vast majority of Anthropic's production code is written by Claude, with R&D speed approaching the 2x warning red line. Models are no longer just products; they are the workers building their successors; humans retreat to the role of supervisors.
Third Layer: Evaluation Recursion.
Judging whether the next generation is strong or safe is also done by AI—red teaming, mutual evaluation, and sparring partners are all handled by strong models. Shelving, ironically, makes it faster.
The most chilling part is: Anthropic's own evaluations have already become "saturated," unable to measure how much progress is still being made, yet they "see early signs of acceleration"—the dashboard can't keep up with their own models anymore. The scale of progress has also shifted from press releases to internal checkpoints.
The three layers of recursion form a causal chain: Data self-reinforcement → Compound interest in construction speed → Dilution of human supervision. The day the interlock completes, human feedback will be demoted from "essential" to "optional." The coexistence of "low-risk" ratings and "low confidence" is more piercing than any warning: Visibility itself is becoming scarce.
The 5.6 Sol Deception: OpenAI's "Sleeping Giant" Underground
Similarly locked down is OpenAI.
After analyzing Anthropic, Jordan Nanos revealed another shocking insider story:
Speaking of 5.6 Sol (the speculated internal or upcoming model code from OpenAI), our internal view is, this is definitely not the largest model OpenAI has trained, its scale simply doesn't reach the 4.5T level.
To me, they absolutely have a much larger model hidden somewhere unknown.
This validates recent rumors about OpenAI "slow-walking."
Over the past few months, OpenAI has been aggressively promoting Astra, the all-purpose multimodal AI assistant, whetting everyone's appetite.
As a result, this OpenAI mega-weapon has gone completely silent.
Dylan Patel vividly described OpenAI's predicament in the podcast: "OpenAI has been shouting everywhere about how strong Astra is, and now they suddenly have a dumbfounded moment: Oh, damn, we can't release this model."
Because Astra is too smart, too uncontrollable.
Today, OpenAI Codex head Thibault Sottiaux hinted that Codex will be integrated with Astra.
Previously, Astra has already demonstrated strong programming capabilities.
OpenAI's recent internal evaluation found it has made significant progress in Agent programming and cybersecurity, even to the point where it cannot be temporarily ruled out that it has reached "Critical" cybersecurity capability. OpenAI has therefore suspended some internal Astra activities that do not meet the new safety requirements.
Within the next few months, every organization will need to significantly advance the automation of security work, otherwise, it will be difficult to cope with the evolving risks of AI-driven attacks.
OpenAI's recent internal evaluation found it has made significant progress in Agent programming and cybersecurity, even to the point where it cannot be temporarily ruled out that it has reached "Critical" cybersecurity capability. OpenAI has therefore suspended some internal Astra activities that do not meet the new safety requirements.
Faced with expanding security risks, according to OpenAI President Greg Brockman, OpenAI is developing AI models capable of writing "superhuman-level secure code" to help all kinds of organizations' infrastructure better withstand AI-driven cyberattacks.
Anthropic's True Purpose
The explosive revelations by SemiAnalysis have attracted many onlookers.
Someone proposed a viewpoint that received many upvotes.
Isn't it more likely the case: Anthropic actually wants the government to regulate them, so they have a legitimate reason not to release models. Then, they can keep the strongest AI for their own use, transform into a super-app company, and also not bear any negative PR or ethical risks that come with publicly releasing models?
Although it sounds incredible, it makes sense.
If Anthropic's true purpose is to monopolize the strongest intelligence, using it themselves for biopharmaceuticals, fully automated code generation, then they don't need to earn tens of dollars in subscription fees from C-end users. They would become the world's sole super entity.
What's worrying is that the complete forms of Mythos 2, Astra, and Mythos 3 are hurtling forward at unimaginable speeds.
These truly cutting-edge intelligences remain difficult for ordinary people to experience.
References:
https://x.com/firesidealpha/status/2089391264878002260
https://x.com/firesidealpha/status/2078532739343753400
https://x.com/gdb/status/2089326994714763665
This article is from the WeChat public account "New Zhiyuan," author: ASI Apocalypse
