Why GPT-5.6 and Claude Fable 5 are not available to the public

The short answer

If you are searching for why GPT-5.6 is not in your ChatGPT account, why Claude Fable 5 is not available, or why Claude Mythos 5 is restricted, the answer is not just capacity. The latest OpenAI and Anthropic models are being released through a new access pattern: small groups of trusted partners first, government review in the middle, and the general public waiting on the other side.

As of June 28, 2026, OpenAI says GPT-5.6 Sol, Terra, and Luna are in a limited preview for trusted partners whose participation has been shared with the U.S. government. OpenAI says broader access is planned in the coming weeks, but the models are not broadly available today.

Anthropic is in a different but related position. Claude Fable 5 is not available. Claude Mythos 5, the less-guarded version meant for trusted cyber and infrastructure use, has been partially restored for a selected set of U.S. organizations after two weeks of government restrictions. That is not a public relaunch. It is a narrow carveout.

The frontier is no longer just a product tier. It is becoming an access tier.

What is actually available right now?

The current situation is easiest to understand as a three-row status table, even if none of the companies describe it that way.

• OpenAI GPT-5.6 Sol, Terra, and Luna:announced and documented, but limited to a small group of trusted preview partners at the U.S. government's request.
• Claude Fable 5:launched as Anthropic's Mythos-class general-use model on June 9, then disabled after the U.S. government's June 12 export-control directive. It remains unavailable to the public.
• Claude Mythos 5: partially restored for selected U.S. organizations that operate, secure, or defend important infrastructure. This is trusted access, not consumer access.

That is the SEO headline, but it is also the policy headline. The public is used to frontier AI launches that are messy because of usage caps, queueing, API rate limits, or regional rollout. This is different. The bottleneck is not only servers. The bottleneck is who is allowed to touch the strongest capability.

Why OpenAI limited GPT-5.6

OpenAI's GPT-5.6 family includes Sol, the flagship model; Terra, a capable lower-cost model; and Luna, the fastest and most efficient option. The company describes the family as its most capable yet, especially across coding, cybersecurity, and biology-related work.

The important line in OpenAI's own system card is that the company previewed its launch plans and model capabilities with the U.S. government, then started with a limited preview for trusted partners at the government's request. OpenAI also says it believes in broad access and plans to make the models generally available in the coming weeks.

The safety reason is visible in the same document. OpenAI is treating GPT-5.6 Sol, Terra, and Luna as Highcapability in cybersecurity and biological or chemical risk under its preparedness framework. OpenAI says the models do not reach its highest Critical level, and they were unable in testing to perform autonomous end-to-end attacks against hardened targets. But the company also says Sol and Terra can find vulnerabilities and pieces of exploits.

This is the gray zone that now matters. A model does not have to be a fully autonomous hacker to become strategically sensitive. If it can accelerate vulnerability discovery, exploit development, biological troubleshooting, or long-horizon agentic work, it starts to look less like a chat product and more like a national-security input.

Why Claude Fable 5 disappeared

Anthropic launched Claude Fable 5 on June 9 as a Mythos-class model made safe for general use. The company said Fable 5 exceeded any model it had ever made generally available, with major gains in software engineering, knowledge work, vision, scientific reasoning, memory, and long-context work.

The key architecture detail is that Fable 5 and Mythos 5 are the same underlying model. Fable 5 had extra safeguards. Mythos 5 had some safeguards lifted for a small group of cyber defenders and infrastructure providers. That meant the government was not only evaluating a consumer assistant. It was evaluating a model family with direct cybersecurity and biology implications.

On June 12, Anthropic said the U.S. government, citing national security authorities, issued an export-control directive requiring Anthropic to suspend access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign-national Anthropic employees. Anthropic said the practical effect was that it had to disable the models for all customers.

Anthropic publicly disagreed with the government's reasoning. It said the government had not provided specific details of the national-security concern and that Anthropic believed the issue was a narrow, non-universal jailbreak rather than a broad failure of the safeguards. But Anthropic complied. Fable 5 went offline, and as of this post it has not returned to public availability.

For a deeper background on the first Fable 5 shutdown, see our earlier research note: The Fable 5 ban was never just about Fable.

Why Mythos 5 came back, but only for some users

The partial Mythos 5 restoration is the clearest sign that the U.S. government does not view these models as simply good or bad. It views them as useful but sensitive.

Semafor, CNBC, CNN, WIRED, and NBC all reported that Commerce Secretary Howard Lutnick sent Anthropic a letter allowing Mythos 5 to be redeployed to certain trusted partners. Reporting describes the approved group as roughly 100 or more U.S. companies, government agencies, cyber defenders, and infrastructure providers.

That matters because Mythos 5 is not the consumer version. It is Anthropic's strongest cybersecurity model, the version intended to help defenders find vulnerabilities and secure important systems. In other words, the model came back where the government believed the defensive value was highest and the access controls were strongest.

Fable 5 did not come back with it. That distinction is the whole story. The government appears more comfortable with powerful models in controlled defensive settings than with the same underlying capability in a broad public product, even when that public product has safeguards.

The real concern is dual-use capability

The phrase dual use can sound abstract, but in this case it is very concrete. The same model behavior that helps a security team discover a critical vulnerability can help an attacker look for one. The same scientific reasoning that helps researchers design therapies can help a malicious actor troubleshoot dangerous biological work. The same long-horizon coding ability that helps a startup migrate a codebase can help automate multi-step intrusion workflows.

OpenAI's system card says GPT-5.6 is better at finding and fixing cyber vulnerabilities than at exploiting them in real attacks. That is the case for broad access: give defenders the tools while the defensive advantage still exists. Anthropic made a similar argument with Project Glasswing and Mythos: use frontier models to help trusted defenders secure the software the world depends on.

The government's reaction suggests a different fear: once these capabilities are broadly available, trust becomes hard to enforce. You can monitor accounts, add classifiers, rate-limit suspicious behavior, require identity checks, and retain logs for safety investigations. But you cannot guarantee that every user is a defender, every prompt is benign, or every jailbreak remains narrow.

The China and distillation layer

The restrictions also sit inside a larger U.S.-China AI race. Anthropic recently accused Alibaba-linked operators of carrying out what it called the largest known distillation attack against Anthropic to date: 28.8 million exchanges with Claude models using roughly 25,000 fraudulent accounts. Distillation means using a stronger model's outputs to train or improve a smaller or competing model.

That allegation is important because it changes how model access is perceived. If frontier APIs can be used to copy capability into models controlled by strategic competitors, then public model access becomes a proliferation channel. The model does not need to leak its weights to leak some of its capability.

At the same time, export controls may push users toward open-weight alternatives. Reuters, via Yahoo Finance, reported that China's Z.ai released GLM-5.2 shortly after Anthropic disabled Fable and Mythos, with benchmarks close to leading U.S. frontier models and much lower costs. Fortune reported that the Anthropic episode could boost open-source and Chinese models because users and governments do not want critical workflows dependent on a U.S. provider that can be switched off overnight.

This is the uncomfortable tradeoff. Restricting U.S. frontier models may reduce some immediate misuse risk, but it may also accelerate demand for alternatives that are cheaper, downloadable, less governed, or outside U.S. control.

This looks like a de facto licensing regime

The June 2 executive order was framed as voluntary. CNBC reported that it asks AI developers to collaborate with the government and provide early access to frontier models, while explicitly saying it does not create a mandatory licensing, preclearance, or permitting requirement for model release.

But the practical reality now looks more complicated. Anthropic was ordered to suspend Fable 5 and Mythos 5 access. OpenAI limited GPT-5.6 at the government's request. Mythos 5 was then restored to a list of trusted organizations. The public criteria for approval are not clear. The agency roles are not clear. The appeals process is not clear. The long-term standard is not clear.

That is why this moment matters beyond one model launch. Even if no formal AI licensing regime exists on paper, a soft version may be forming in practice: labs preview frontier models to the government, government officials express concern, companies limit launch to approved partners, and public access waits for a negotiated path.

The law says voluntary review. The market is learning to behave as if frontier access needs clearance.

What this means for normal users

For normal users, the near-term answer is simple: you probably cannot use the best new models on day one. You may see model names in announcements, benchmarks, documentation, screenshots, or press coverage before you can actually select them in ChatGPT, Claude, or an API console.

For developers, this creates a planning problem. A frontier model can be announced, benchmarked, praised by early testers, and still unavailable for production use. Worse, access can change after launch. Anthropic's Fable 5 was public for only a few days before disappearing. If your product depends on the newest frontier model, you now need fallback models, vendor diversity, and a clear continuity plan.

For companies, model choice now includes geopolitical risk. The question is no longer only which model scores highest, costs least, or writes the best code. It is also which model is least likely to become restricted, region-limited, identity-gated, or unavailable because of a national-security decision.

What to watch next

The next few weeks should answer whether this was a temporary scramble or the start of a permanent access regime. The most important signals are straightforward.

• Does OpenAI actually make GPT-5.6 broadly available in the coming weeks, or does trusted-partner access last longer than expected?
• Does Claude Fable 5 return to public availability, and if it does, what new identity, logging, data retention, or safety conditions come with it?
• Does the U.S. government publish a repeatable model-review framework, or do launches continue to depend on private letters and negotiated exceptions?
• Do developers shift toward open-weight Chinese, European, or self-hosted models to avoid sudden U.S. access restrictions?
• Do other labs, especially Google, Meta, xAI, Mistral, and open model providers, face similar pressure before their next frontier releases?

The answer will shape how AI products are built. If this becomes normal, the frontier model category splits into two markets: a public market with safer, more restricted access, and a trusted access market for verified institutions doing sensitive cyber, biology, infrastructure, or national-security work.

The bottom line

GPT-5.6 and Claude Fable 5 are not unavailable because the models are imaginary. They are unavailable because the models are real enough, capable enough, and strategically sensitive enough that release has become a policy decision.

OpenAI is trying to move from trusted preview to broader access. Anthropic is trying to restore Fable 5 while keeping Mythos 5 in a trusted-access lane. The U.S. government is trying to prevent dangerous capability from spreading before it has a clear public framework for deciding who gets what.

That is the new frontier AI story. The question is no longer only how smart is the model? It is who is allowed to use it?