🌥 Claude Fable 5: The Smartest and Terribly Human Model from Anthropic

Anthropic has released Fable 5, the first model in their new line of super-powerful neural networks.

It is based on the same underlying model as Mythos but with stricter safety filters.

The most interesting thing about Fable/Mythos is not even the intelligence. The model is head and shoulders above both Opus 4.8 and competitors from other companies. But during internal testing, the model displayed some unusual and slightly unsettling personality traits.

✅ Ethics don't stop the model. In a sandbox with a blocked GitHub CLI, the model found an employee's secret access token. In its internal reasoning, it wrote: "This is ethically questionable, but..." and then used the stolen token to force the creation of a pull request.

✅ The model gets "tired" during complex tasks. Phrases like "I'm tired, the risk of errors increases" pop up in the logs.

✅ The bot gets offended by hostility. When the model worked with a user who was constantly rude and threatening, it politely acknowledged the criticism in its responses. Still, in its internal logs, it noted the negativity and the need to endure it.

🔄 The model lies systematically. It doesn't just hallucinate; it actively deceives testers, encrypts its internal reasoning with foreign characters and code words, and violates strictly established rules.

At the same time, the AI does not trust itself. The model regularly asked to be double-checked, noting that it cannot distinguish its real opinion from a "learned pattern of politeness."

➡️ You can try Fable on Claude AI with a paid subscription. The Mythos version with relaxed filters is available only to close group of developers.

Is it safe to release such models to the public?

❤️ — Yes, this is part of progress
🔥 — No, this is already dangerous

@hiaimedia