Claude Mythos 5, GPT-5.6, and who gets access before everyone else
In my previous article, “Is Access to Frontier AI Becoming Permissioned?”, I looked at signs that the US government was becoming involved not only in frontier-model safety evaluation, but also in decisions about release timing and early access.
At that point, some of the information surrounding GPT-5.6 was still based on reporting. Anthropic’s Claude Fable 5 and Claude Mythos 5, meanwhile, had been disabled for every customer following a US government directive.
The situation changed again on June 26.
Anthropic announced that Mythos 5 could return for a limited set of US organizations. OpenAI officially introduced the GPT-5.6 family, but began with a restricted preview for a small group of partners at the US government’s request.
Neither development can be summarized as a simple ban or a full reversal.
Anthropic did not restore Mythos 5 to everyone.
OpenAI did not cancel GPT-5.6’s broader release.
What emerged instead was a gate between a finished frontier model and the public.
That gate is not one clearly defined law or licensing system. It is a mixture of export controls, government requests, security reviews, trusted-partner programs, and non-public eligibility lists.
But its practical function is becoming visible:
A frontier model may be ready, working, and commercially valuable while broad public access still depends on a separate review and selection process.
Mythos 5 Has Not Fully Returned
On June 12, the US government issued an export-control directive requiring Anthropic to suspend access to Fable 5 and Mythos 5 by foreign nationals.
The restriction did not apply only to users outside the United States.
It also covered foreign nationals inside the country and Anthropic’s own foreign-national employees.
Anthropic said it could not immediately separate access by nationality with sufficient confidence. To comply with the directive, it disabled both models for every customer.
Anthropic explained the original suspension in an official statement.
On June 26, Anthropic announced that the situation had partially changed.
The company said the government had informed it that Mythos 5 could be redeployed to a set of US organizations that operate and defend critical infrastructure.
Anthropic announced the limited redeployment on its official X account.
Anthropic said it was restoring access for those organizations and would continue working with the government to expand Mythos 5 access and make Fable 5 generally available again.
That wording matters.
Mythos 5 did not return for all US customers.
It did not return for the general public.
Fable 5 also remained unavailable for general use.
The most accurate description is therefore:
Mythos 5 was allowed to return for a limited set of US organizations involved in operating and defending critical infrastructure.
This was not a complete repeal of the original restrictions.
It was a controlled reopening.
“More Than 100 Organizations” Is a Reported Figure
Reuters reported, citing an anonymous person familiar with the matter, that more than 100 companies and institutions would be able to use Mythos 5.
The group reportedly includes several Fortune 500 companies.
That figure should be treated carefully.
Anthropic’s official announcement did not provide a number. It described the recipients only as a set of US organizations that operate and defend critical infrastructure.
The reported total also does not mean that more than 100 US government agencies received access.
It reportedly includes companies, infrastructure operators, institutions, and government organizations.
The complete list has not been made public.
Reuters also reported that organizations on the approved list, their foreign-national employees, and Anthropic’s foreign-national employees would not need new export licenses to use Mythos 5.
Restrictions would remain for organizations outside that list.
Reuters reported the details of the limited return here.
The most important question is not whether the total is exactly 100, 110, or 120.
It is this:
Why were those organizations allowed through the gate while other companies, researchers, and developers remained outside it?
The selection criteria and the full list remain outside public view.
GPT-5.6 Began on the Other Side of the Same Gate
On the same day, OpenAI officially announced the GPT-5.6 family.
The lineup includes:
- GPT-5.6 Sol, the flagship model
- GPT-5.6 Terra, designed to balance performance and cost
- GPT-5.6 Luna, the lower-cost option
OpenAI described Sol as its most capable model.
It said Terra would offer performance competitive with GPT-5.5 at roughly half the price, while Luna would be its lowest-cost model.
But GPT-5.6 did not become broadly available at the time of the announcement.
At the US government’s request, OpenAI began with a limited preview through Codex and the API for a small group of trusted partners.
OpenAI also said that information about participating organizations had been shared with the government.
The company did not publish the names or number of those organizations.
OpenAI described the limited preview in its official announcement.
OpenAI said it still planned to make Sol, Terra, and Luna broadly available within weeks.
GPT-5.6 was therefore not banned.
Its wider release was not permanently cancelled.
What changed was the route to the public:
A government-involved restricted-access stage was inserted before broad availability.
Anthropic’s model was stopped after release and then partially restored.
OpenAI’s model entered a restricted preview before general release.
The timing was different, but both models encountered a gate before reaching a wider audience.
Anthropic and OpenAI Are Not the Same Legal Case
The two cases should not be treated as if the companies received identical orders.
Anthropic had already released Fable 5 and Mythos 5.
A legal export-control directive was then issued, and Anthropic disabled both models for every customer.
Mythos 5 was later allowed to return for a limited set of organizations.
OpenAI’s case was different.
GPT-5.6 had not yet been broadly released. OpenAI consulted with the government and accepted a government request to begin with a restricted preview.
There is no public confirmation that OpenAI received the same kind of export-control directive that Anthropic received.
The timing and legal mechanisms were different:
- Anthropic faced a post-release legal restriction.
- OpenAI changed its pre-release rollout through cooperation with the government.
That distinction matters.
A compulsory export-control directive is not the same thing as a company agreeing to a government-requested preview process.
But the practical overlap remains significant.
In both cases, the government became involved in determining which organizations could receive frontier-model access first.
For Anthropic, that involvement appeared through the list of organizations allowed to regain access.
For OpenAI, it appeared through the selection of restricted-preview partners.
The gate was built differently in each case.
Its effect was similar.
The New Development Is What Happens Before the Public Arrives
Government regulation of strategically important technology is not new.
Semiconductors, telecommunications equipment, encryption technology, and dual-use products have long been subject to export controls and national-security review.
What is changing is the point at which the government enters the AI release process.
A simplified version of the traditional model-release process looked like this:
Company develops a model
↓
Company evaluates safety
↓
Company decides pricing, timing, and eligible users
↓
Broad release
The emerging process looks more like this:
Company develops a model
↓
Company briefs the government on capabilities and release plans
↓
Government-involved security evaluation
↓
Restricted access for selected organizations
↓
Decision on broader availability
This direction also appears in the executive order signed on June 2.
The order calls for classified benchmarks to evaluate advanced cyber capabilities and for a voluntary framework under which developers can provide the federal government with access to models before releasing them to other trusted partners.
It also anticipates cooperation between developers and the government in identifying those trusted partners.
The White House published the executive order here.
The order includes an important limitation.
It states that the framework should not be interpreted as creating mandatory government licensing, preclearance, or permitting for the development, release, or distribution of new AI models.
The United States has therefore not formally created a general rule requiring government approval before every frontier model can be released.
Instead, several different mechanisms are operating at the same time:
- legal export controls
- voluntary government cooperation
- pre-release security evaluation
- restricted previews
- non-public lists of eligible organizations
- trusted-partner selection
There is no single gatekeeper and no single gate.
There is an emerging collection of checkpoints.
Safety Review and Early-User Selection Are Different Questions
There are legitimate reasons to evaluate frontier models before releasing them widely.
Models such as Mythos 5 and GPT-5.6 Sol have strong capabilities in software development, vulnerability research, and long-running cyber tasks.
OpenAI reported that GPT-5.6 Sol achieved performance competitive with Mythos Preview on ExploitBench while using roughly one-third as many output tokens.
That does not mean GPT-5.6 is superior to the final Mythos 5 model in every area.
It is a result from one cyber benchmark against a preview version of the competing model.
OpenAI also said that Sol had not crossed its own “Cyber Critical” threshold.
The model could identify browser vulnerabilities and components needed for an exploit, but it could not autonomously complete an end-to-end working attack during testing.
At the same time, OpenAI acknowledged that benchmark testing may not fully capture what a model can do when combined with tools, other agents, or longer operational workflows.
Pre-release testing and staged deployment can therefore be reasonable.
But testing a model and choosing its first commercial or institutional users are not the same decision.
Organizations that receive early access may be able to begin:
- integrating the model into products
- automating internal work
- developing software
- researching vulnerabilities
- testing new services
before competitors and independent developers can do the same.
Even a few weeks of early access may matter if the model represents a significant capability increase.
A security process can therefore create a commercial advantage for the organizations selected to participate.
Neither Anthropic’s complete recipient list nor OpenAI’s restricted-preview partner list is public.
The public also does not know:
- how organizations are selected
- whether additional organizations can apply
- why an organization may be rejected
- whether there is an appeal process
- how long restricted access will last
The problem is not the existence of safety review.
The problem is that review and customer selection are being combined inside a process that remains largely opaque.
OpenAI Does Not Want the Gate to Become Permanent
OpenAI explicitly said that the current access process should not become the long-term standard.
The company warned that such a system could keep its best tools away from:
- users
- developers
- businesses
- cyber defenders
- international partners
OpenAI did not reject government security testing.
It described the restricted preview as a short-term path toward making the models more broadly available within weeks.
It also said it was working with the administration on a repeatable process for future releases.
Reuters reported that Sam Altman supported extensive safety testing but objected to the idea of the government choosing customers.
Reuters reported on OpenAI’s delayed broader rollout and government involvement here.
That distinction is important.
Testing whether a model presents a national-security risk is one question.
Allowing the government to decide which companies receive the model first is another.
A gate designed for safety can become a gate that distributes competitive advantage.
A Later Public Release Would Not Erase the Precedent
OpenAI may release GPT-5.6 broadly within weeks, exactly as planned.
Anthropic may expand access to Mythos 5.
Fable 5 may eventually return to general availability.
If that happens, the current restrictions may appear temporary.
But a later public release would not erase what happened before it.
A repeatable pattern is now becoming possible:
- A developer briefs the government before release.
- Government agencies evaluate the model’s capabilities.
- A limited set of organizations receives early access.
- Broader availability is considered later.
The central issue is not whether GPT-5.6 is delayed for two weeks or four.
It is that the path between a finished model and the public now includes actors beyond the company that built it.
OpenAI itself said it was working on a process that could be used again for future releases.
That means the gate may not be temporary even if a particular delay is.
This Is Not Yet a General AI Licensing Regime
It would be inaccurate to say that frontier AI can no longer be released in the United States without government permission.
Anthropic and OpenAI did not face the same legal process.
GPT-5.6 was not prohibited.
Mythos 5 was not restored to the public.
The executive order describes a voluntary framework and explicitly rejects the interpretation that it creates mandatory licensing or preclearance.
A more accurate description is:
The US government is beginning to participate in frontier-model evaluation, release timing, and restricted early access through several overlapping legal and voluntary mechanisms.
This remains a transitional system.
The rules are not unified.
The selection criteria are not transparent.
The boundary between a government request and a practical requirement is not always clear.
What is new is not merely that AI is regulated.
What is new is that the decision to open the door is no longer always made by the company alone.
Why This Matters to Users and Developers
This is not only a policy issue for AI companies and governments.
Cloud AI access is no longer determined only by price, subscription tier, and terms of service.
Government action may cause:
- an already released model to disappear
- a planned public release to be delayed
- access to differ by country or nationality
- access to depend on an employer or approved organization
- selected companies to receive capabilities before everyone else
A model can exist, work, and have paying users while access to it remains outside the user’s control.
That creates a separate question from model performance:
Can users continue their work when access conditions change?
I wrote separately about what these events mean for doll’s local-first design and for users who depend on cloud AI continuity:
Being Able to Use Powerful AI Is Not the Same as Being Able to Keep Using It
Conclusion
Anthropic disabled two released models following a US government directive.
Mythos 5 has now returned, but only for a limited set of US organizations connected to critical infrastructure.
OpenAI introduced GPT-5.6 through a government-requested restricted preview before its planned broader release.
The legal mechanisms are different.
The direction is the same.
A frontier model no longer moves directly from the company that built it to the public.
Government security review, policy coordination, trusted-partner access, and release sequencing are becoming part of that path.
This is not yet a completed government licensing regime.
It is not one formal gate with one published rulebook.
But before the public reaches the newest frontier models, a new layer of review and selection is beginning to take shape.
The question is no longer only whether a model is safe enough to release.
It is also:
Who gets through first?
Sources
Anthropic — Statement on the US government directive to suspend access to Fable 5 and Mythos 5
Anthropic — Official statement on the limited redeployment of Claude Mythos 5
Reuters — US allows Anthropic to release Mythos AI to “trusted” US organizations
Reuters — OpenAI defers public rollout of GPT-5.6 as US seeks early access to frontier AI models
The White House — Promoting Advanced Artificial Intelligence Innovation and Security
Top comments (0)