Anthropic Released a Powerful AI with Safety Limits

Anthropic released a new AI model last Tuesday. They called it Claude Fable 5. The company said it was their most powerful model. But they also added special limits. The model will not answer questions about cybersecurity, biology, or chemistry. The company worried that bad people could use the answers to cause harm. Fable 5 used the same technology as another model called Mythos 5. Only a small group of trusted cybersecurity experts could use Mythos 5. The public could use Fable 5, but the system redirected dangerous questions to an older model. That older model was Opus 4.8. The safety system sometimes blocked harmless requests. This happened in less than five percent of tests. Anthropic said these false positives were acceptable. They wanted to prevent serious harm. The company tested Fable 5 for more than 1,000 hours. External teams tried to break the safety rules. They did not find any universal way to do it. The company worried most about the model's ability to perform multi-part cyberattacks. The UK's AI Security Institute tested the model. They found it was similar to another AI from OpenAI called GPT-5.5. Companies could use Fable 5 through an API. They had to pay $10 for every million input tokens. They had to pay $50 for every million output tokens. These prices were much higher than those for GPT-5.5.