Pre-AGI and the Public AI Ceiling
Why Frontier Models Are Moving Behind Access Controls
Something important has changed in artificial intelligence.
And strangely, the biggest change may not be that AI models are becoming more intelligent.
It may be that access to intelligence itself is beginning to fragment.
Look at the current generation of frontier AI models: GPT-6 Astra, Claude Fable 5, Kimi K3, GLM-5.3 and Qwen3.8-Max.
These systems can reason, code, research, operate tools, process enormous contexts and perform increasingly complex agentic work. They are no longer simple chatbots.
I call this generation Pre-AGI.
Not because “Pre-AGI” is an officially defined technical category. It isn't.
I use the term because these models appear to occupy a fascinating space between traditional generative AI and whatever ultimately deserves to be called Artificial General Intelligence.
And I increasingly believe we are approaching a ceiling.
But not a ceiling on AI development.
A ceiling on unrestricted public access to frontier AI capabilities.
The Public AI Ceiling Is Different From the Capability Ceiling
This distinction matters.
AI laboratories are not suddenly going to stop training stronger models.
OpenAI, Anthropic, Google, Alibaba, Moonshot AI, Z.ai and other frontier laboratories will continue pushing model intelligence, reasoning, coding, science, cybersecurity and autonomous-agent capabilities forward.
The real change appears to be happening somewhere else.
Instead of releasing every capability equally to everyone, frontier AI companies are beginning to create different layers of access.
The general public receives one version.
Verified researchers may receive another.
Cybersecurity professionals may receive fewer restrictions.
Critical-infrastructure operators may receive more specialized capabilities.
Governments and national-security organizations may operate models under entirely different access arrangements.
And the laboratories themselves will naturally have access to internal checkpoints and experimental systems that never reach ordinary users.
That structure is no longer speculation.
We can already see it appearing.
Claude Fable 5 and Mythos 5 Show What Restricted Frontier AI Looks Like
Anthropic has provided one of the clearest examples.
Claude Fable 5 and Claude Mythos 5 use the same underlying Mythos-class model, but they do not provide identical access to its capabilities.
Fable 5 is designed for broad public use.
Mythos 5 provides selected users with fewer restrictions in certain high-risk domains.
Anthropic has publicly explained that Fable 5 uses separate safety classifiers covering areas including cybersecurity, biology, chemistry and model distillation.
When those classifiers trigger, some requests can be routed away from the underlying Fable capability and handled instead by Claude Opus 4.8.
Anthropic has said that this fallback occurs in fewer than 5% of Fable sessions overall, although specific high-risk workloads may encounter it considerably more often.
This distinction is extremely important.
It means that a user may technically have access to a frontier model while not having unrestricted access to every capability contained within that model.
Mythos 5 demonstrates the other side of that equation.
Anthropic initially provided Mythos access to selected cybersecurity defenders and critical-infrastructure partners through Project Glasswing, developed in collaboration with the U.S. government.
The company describes Mythos 5 as the same underlying model as Fable 5, but with certain safeguards removed for authorized users.
That is not ordinary product segmentation.
That is capability segmentation.
One correction is important here.
Calling Mythos 5 “classified” would go further than the available evidence supports. There is no publicly established evidence that Mythos itself carries a formal national-security classification.
The more accurate description is:
restricted, trusted-access frontier AI.
And that distinction is actually more interesting.
Governments Are Already Treating Frontier AI as a National-Security Issue
The relationship between frontier AI and governments is becoming increasingly direct.
In June 2026, the U.S. government temporarily imposed restrictions affecting access to Claude Fable 5 and Mythos 5 for foreign nationals under national-security export-control authorities.
Those restrictions were later lifted.
But the precedent matters.
The episode demonstrated that access to high-capability AI models can move from being a normal commercial decision to becoming part of national-security policy.
That is a major transition.
For decades, governments controlled access to advanced weapons systems, cryptography, semiconductor manufacturing technology, aerospace systems and other strategically important technologies.
Frontier artificial intelligence is beginning to enter the same conversation.
GPT-6 Astra Makes the Access Divide Even Clearer
OpenAI's GPT-6 Astra provides another powerful signal.
OpenAI describes Astra as the first model to reach the Critical cybersecurity capability threshold under its Preparedness Framework.
According to OpenAI, with appropriate tools and access, Astra can identify previously unknown vulnerabilities and develop methods to exploit hardened systems.
That level of capability creates an obvious dilemma.
The same intelligence that can defend critical infrastructure can potentially attack it.
So OpenAI has not treated every Astra capability as something that must automatically be available to every user.
The company has explicitly stated that access to Astra's most advanced cybersecurity capabilities will be more limited.
Initial access has been offered to selected testers, with expansion occurring through trusted-access initiatives such as Daybreak.
OpenAI has also described additional safety systems surrounding Astra, including stronger refusal boundaries, monitoring, classifiers and mechanisms capable of interrupting potentially unauthorized activity.
Calling Astra “filtered” is therefore directionally correct.
But a better term may be:
capability-gated.
The intelligence exists.
Access to certain expressions of that intelligence is controlled.
OpenAI Already Operates Multiple Levels of AI Access
Astra is not an isolated case.
OpenAI has already experimented with different access tiers through its cybersecurity programs.
Models such as GPT-5.5-Cyber were made available through limited programs designed for vetted defenders and organizations responsible for critical infrastructure.
Verified users can encounter fewer restrictions when conducting legitimate defensive cybersecurity work.
More sensitive capabilities require stronger verification and monitoring.
That creates something very different from the AI model market we became accustomed to between 2022 and 2025.
The old model was simple:
A company released a model.
Everyone received roughly the same model.
The new structure increasingly looks like this:
same intelligence — different permissions.
That may become one of the defining characteristics of frontier AI.
Kimi K3, GLM-5.3 and Qwen3.8-Max Complicate the Picture
There is, however, an important counterforce.
Powerful open and open-weight models make a universal public AI ceiling extremely difficult to enforce.
Kimi K3 is a 2.8-trillion-parameter model with a one-million-token context window and openly released weights.
Alibaba's Qwen3.8-Max represents another enormous frontier system, built for long-context reasoning, multimodal intelligence and professional workloads.
GLM-5.3 similarly pushes open-model capabilities deeper into coding, cybersecurity and agentic work.
This matters because intelligence cannot be perfectly controlled once sufficiently powerful model weights are widely distributed.
There is no global switch capable of shutting down frontier AI development.
Open models will improve.
Distillation will improve.
Hardware will become faster.
Inference will become cheaper.
Agent frameworks will become more sophisticated.
Researchers will combine multiple models into increasingly powerful systems.
So the public AI ceiling I am describing is not absolute.
It is better understood as a managed frontier.
The Future May Have Several AI Frontiers
For most of the AI era so far, we talked about the frontier.
There was one frontier.
The strongest publicly visible model defined it.
That concept may now be becoming obsolete.
Instead, we may soon have several different AI frontiers operating simultaneously.
There will be a public AI frontier, accessible to millions of users.
There will be a trusted-access frontier, available to verified researchers, cybersecurity professionals and selected organizations.
There will be an enterprise and critical-infrastructure frontier, where models operate with deeper integrations, larger toolsets and stronger permissions.
There will likely be a national-security AI frontier, where defense, intelligence, cybersecurity and government institutions use capabilities that are not ordinarily exposed to the public.
And above all of them will remain the internal research frontier.
This final layer may matter the most.
AI laboratories naturally evaluate systems before releasing them.
They possess experimental checkpoints.
They test unfinished successors.
They connect models to specialized tools.
They run internal agent systems.
They explore capabilities long before those capabilities become commercial products.
The frontier visible to the public therefore does not necessarily represent the frontier of what exists.
This Does Not Mean Secret AGI Has Been Proven
This is where the conversation requires discipline.
There is currently no public evidence proving that OpenAI, Anthropic, Google or any government already possesses a fully developed secret AGI operating behind closed doors.
Claiming that as established fact would weaken the argument.
The evidence supports something more defensible — and arguably more important.
The infrastructure required to restrict future frontier intelligence is already being built.
Trusted-access programs exist.
Capability-specific safeguards exist.
Model routing exists.
Identity verification exists.
Specialized cyber variants exist.
Government involvement exists.
Critical-infrastructure partnerships exist.
Different permission levels exist.
The machinery for a stratified AI ecosystem is appearing before AGI itself has even been publicly established.
The Top 100 Organizations May Live in a Different AI World
My own forecast goes further.
I suspect that once AI systems become substantially more capable than today's Pre-AGI models, unrestricted frontier access will become increasingly concentrated.
Defense departments will obviously have strong incentives to obtain the deepest capabilities.
National cybersecurity agencies will want them.
Intelligence organizations will want them.
Frontier laboratories themselves will possess them.
The world's largest hyperscalers, semiconductor companies, infrastructure providers and technology corporations will have both the resources and strategic importance necessary to justify access.
My estimate that perhaps the top 100 organizations globally may initially operate at this higher frontier is not a documented fact.
It is a forecast.
But the institutional architecture required for such a system is already becoming visible.
And once frontier intelligence becomes capable of autonomously discovering zero-day vulnerabilities, conducting advanced scientific research, designing biological systems, operating industrial infrastructure or coordinating thousands of autonomous agents, unrestricted distribution becomes much harder to justify.
The Next AI Race May Become Invisible
The AI race from 2022 onward was extraordinarily public.
We watched GPT-4 arrive.
Then stronger Claude models.
Then Gemini.
Then DeepSeek.
Then Qwen.
Then Kimi.
Every few months benchmarks moved.
Context windows grew.
Coding scores improved.
Reasoning became deeper.
Millions of people could immediately test the progress themselves.
The next phase may not look like that.
Imagine a public model improving from capability level 90 to 95.
Everyone sees it.
Benchmarks appear.
Developers test it.
YouTube fills with demonstrations.
Now imagine an internal system reaching capability level 110, but the laboratory decides it cannot responsibly release the full tool access, autonomous runtime or cyber capability.
There may be no public benchmark.
No API.
No consumer launch.
No leaderboard.
The system exists, but the public cannot directly measure it.
That changes the nature of technological competition.
The most important AI race could gradually disappear behind secured infrastructure.
“The Most Powerful AI Available” May Become a Meaningless Phrase
For years we asked:
Which AI model is the most powerful?
Soon that question may become incomplete.
We may instead have to ask:
Which version?
With which tools?
With which permissions?
With which safety layer?
Under which trusted-access program?
Available to which organization?
Those questions will matter because intelligence may no longer be distributed as a single homogeneous product.
The same underlying model could behave very differently depending on who is using it and what level of access they have been granted.
That is already happening with Claude Fable 5 and Mythos 5.
We can already see elements of it in OpenAI's Trusted Access programs.
And Astra appears to push the architecture further.
Welcome to the Pre-AGI Access Era
Astra, Fable 5, Kimi K3, GLM-5.3 and Qwen3.8-Max may eventually look primitive compared with what follows.
That is not the point.
Their historical importance may be something else entirely.
They may represent the generation in which society first discovered that the frontier of artificial intelligence could no longer be distributed uniformly.
Development will continue.
Models will become stronger.
Agents will become more autonomous.
Scientific capabilities will grow.
Open models will keep challenging closed systems.
But above a certain capability threshold, the question may no longer be simply:
What can the model do?
The defining question of the Pre-AGI era may become:
Who is permitted to let it do it?
That is the public AI ceiling I believe we are beginning to see.
The intelligence race is not stopping.
It is becoming layered.
And eventually, the most consequential layer may no longer be public.
© 2025 Kohenoor Technologies. All rights reserved.
Support: info@kohenoor.tech | Privacy Policy | Terms
Kohenoor Technologies is a global technology ecosystem building KENOS; the Operating System for the Intelligent Era. KENOS unifies Action AI, Hybrid Finance (HyFi), blockchain and digital-asset infrastructure, cybersecurity, enterprise intelligence, and Education 3.0.
Powered by KAI and Fueled by KEN, the intelligent engine helps enterprises, institutions, and individuals adopt AI, modernize digital and financial infrastructure, strengthen security, and build future-ready skills for the AI-driven global economy.