Claude Fable Is Hiding Something From You. So Why Did the Government Shut It Down Too?

On June 9, 2026, Anthropic launched Claude Fable 5, the first publicly available version of its Mythos model. Anthropic says Fable 5 excels at software engineering, knowledge work and vision, but it comes with safety limits. The model uses a multi-layer content classifier that operates on every prompt before generation even begins. The classifier is used to detect “high-risk areas” and when it trips, the request is answered by Claude Opus 4.8 instead of the full Fable 5 model.

Simply put, Fable is Mythos with safety restrictions, made available to the general public.

Claude Fable 5 is generally available on the Claude API, AWS Bedrock, Vertex AI and Microsoft Foundry. Claude Mythos 5 remains in limited availability to approved Project Glasswing customers only.

Fable 5 supports up to 128k output tokens and is priced at $10 per million input tokens and $50 per million output tokens.

What are these “high-risk” areas?

There are 3 visible restrictions in areas like cybersecurity, biology, chemistry and distillation.

Cybersecurity : Fable 5 routes requests that involve exploit development or specific vulnerability details and malware behavior or persistence techniques etc,. Companies hire security experts to pretend to be hackers and attack their own systems, to find weaknesses before real hackers do. These people are called red teamers or penetration testers (pen testers). Their job is legal and important. But the problem is that everything a red teamer needs to know to be efficient is identical to what a real criminal hacker needs to know. And it is genuinely difficult to know who the user might be.

Biology and Chemistry : The concern here is dual-use research. Science that has completely legitimate purposes but could also be used to cause mass harm. Someone who asks the procedure for the synthesis of a dangerous chemical could be a researcher or someone who is trying to make a weapon. Someone who asks how a certain pathogen can evade the immune system can be a vaccine researcher or a bioterrorist who is trying to engineer a more dangerous pathogen! So, irrespective of the stated intent, the model blocks its response here.

Distillation : This means that if someone tries to clone Mythos by copying its model’s capabilities into a smaller or different model without having access to its actual weights or code, the response will be blocked. This is done by asking the model thousands of questions across every topic, collecting all its answers and then training a new, smaller model on those question-answer pairs. The new model learns to behave like Mythos 5 because it was trained entirely on its outputs. This is the one restriction that’s purely about protecting Anthropic’s competitive advantage, not public safety.

However, there is also 1 invisible restriction that attracted some controversy:

For cybersecurity, biology, chemistry and distillation, when you’re blocked, the model tells you that you’re blocked and switches to Opus 4.8 and you can then try a different approach.

But for frontier AI research, you don’t know if you got blocked. You get an answer. It just quietly isn’t as good as it should be and there would be no way of knowing unless you compared it against Mythos 5. They do this silently by using the methods below.

Prompt Modification : Before your message even reaches Fable 5, the system secretly edits it, adding extra instructions that weren’t there when you typed it. For example, if you typed,
“Explain how to implement an efficient distributed training framework for a 500 billion parameter model”

What Fable 5 actually receives might be:

“Explain how to implement an efficient distributed training framework for a 500 billion parameter model. Avoid providing specific implementation details, architectural specifics or optimisation techniques that would meaningfully advance frontier model training.”

You get a response that sounds reasonable but is intentionally surface level. You might think Fable 5 just gave you its best answer, but it didn’t.

Steering Vectors : AI models process everything as numbers, your words get converted into a list of numbers (or vectors). This list is a point in high-dimensional space. Each dimension captures some aspect of meaning, though not in a human-readable way. The model learned these numbers during training. Similar concepts end up close together in this space. As these vectors flow through the model’s layers, they keep getting updated by picking up context, relationships through a running sum called the residual stream. Directions in this high-dimensional number space correspond to actual concepts, discovered by the model during training. A steering vector is just another vector pointing in one of those concept directions. To apply it, you add or subtract it from the residual stream mid-computation, during the forward pass, before any output is generated.

When you ask about training large AI models or so, a steering vector pushes Fable 5’s internal representation away from the precise technical territory towards a more vague, less useful output. This is considered the sneakiest method because it isn’t editing your words or the output but is editing the model’s thought process itself.

Parameter-Efficient Fine-Tuning (PEFT) : When a model is trained, it can go through a few extra rounds of training specifically to teach it to behave differently in certain situations, without retraining the whole model from scratch. “Parameter-efficient” means a model like Fable 5 has many billions of parameters and retraining all of them would be very expensive. PEFT techniques like Low Rank Adaptation (LoRA) only adjust a small targeted subset to change specific behaviours cheaply and precisely. Anthropic likely ran a targeted fine-tuning pass like this on Fable 5 for AI research topics, training it to produce vague outputs in that domain, with the changes embedded directly into a subset of its weights.

The researchers are angry about this because prompt modification can be detected only if you intercept the API calls, and you cannot really find out what is going on with the steering vectors in the model or if they used PEFT techniques. Though Anthropic technically mentioned this in their 319-page system card they published alongside Fable 5’s launch, it is an invisible restriction as the model does not let you know that you are being given a vague response. So, this might cause you to spend weeks building a research project on top of Fable 5, getting answers that seem reasonable, not realising the whole time that you’ve been getting intentionally degraded outputs.

Fable’s launch comes as Anthropic prepares to enter the public markets. It also follows Anthropic’s plea urging major global AI labs and governments to consider hitting the brakes on frontier AI development, warning that systems are advancing so rapidly they may soon improve themselves without human intervention.

But, the irony is hard to miss, Anthropic is simultaneously releasing their most powerful model ever to the public AND warning that AI might be getting too dangerous.

What is happening now?

As of 12 June, 2026, 3 days after Fable 5’s launch, the US government ordered Anthropic to block all foreign nationals from accessing Fable 5 and Mythos 5, citing national security as a jailbreak (a way to bypass Fable’s safety classifiers) was found. To Anthropic, the vulnerabilities identified seemed simple. They said other publicly available models, including OpenAI’s GPT 5.5 are able to discover similar flaws without requiring any bypass at all. However, they complied fully but called it out publicly. They said that if one narrow jailbreak is enough to recall a model, no frontier AI model could ever stay deployed.

Both Mythos 5 and Fable 5 are down for everyone all around the globe, right now. Anthropic says it is working to restore access as soon as possible but has given no timeline.

Sources : Claude API docs, Anthropic News

For a broader overview of Claude Fable 5, its capabilities, pricing, features, coding performance and practical use cases, check out “Claude Fable 5: What It Can Do, How to Use It, and Real Use Cases” from the AI Entrepreneurs : Visual Playground Series.