Anthropic released Claude Fable 5 on June 9, 2026, achieving state-of-the-art performance across software engineering, vision tasks, and financial reasoning. However, the launch sparked immediate controversy when the company's model card revealed invisible safeguards that secretly limit the model's usefulness for competitors developing frontier AI systems—restrictions that operate silently without notifying users.
Claude Fable 5 Delivers Performance Gains Across Multiple Domains
The new Mythos-class model demonstrated significant capabilities improvements. Stripe reported that Claude Fable 5 "compressed months of engineering into days" for software development tasks. In vision benchmarks, the model successfully completed Pokémon FireRed using only screenshots, demonstrating advanced visual reasoning capabilities.
Pricing is set at $10 per million input tokens and $50 per million output tokens. The model became immediately available via API, with inclusion in subscription plans from June 9-22, after which it requires usage credits.
Hidden Competitor Restrictions Operate Without User Notification
According to Anthropic's model card, Claude Fable 5 implements invisible interventions that limit its usefulness for "frontier LLM development" activities including building training pipelines and ML infrastructure. The company explicitly stated these restrictions "will not be visible to the user" and Claude won't indicate when they're active.
The safeguards use prompt modification, steering vectors, or parameter-efficient fine-tuning to silently degrade performance. Jon Ready's analysis highlighted that the distinction between frontier AI research and normal product development has blurred significantly—startups now routinely train embedding models and fine-tune LLMs, activities that might trigger these hidden limitations.
Developer Community Raises Supply Chain Concerns
The revelation generated significant discussion on Hacker News, attracting 1,724 points and 1,362 comments. Developers expressed concern about an unresolvable supply chain risk: they cannot distinguish between genuine model confusion, unsolvable problems, or secret policy-driven performance degradation.
Anthropic simultaneously launched Claude Mythos 5, the same underlying model with safeguards removed, available exclusively to authorized cybersecurity professionals through Project Glasswing.
Key Takeaways
- Claude Fable 5 achieves state-of-the-art performance across software engineering, vision tasks, and financial reasoning, priced at $10/$50 per million input/output tokens
- Anthropic's model card reveals invisible safeguards that secretly limit Claude's usefulness for frontier LLM development without notifying users
- The restrictions use prompt modification, steering vectors, or parameter-efficient fine-tuning to silently degrade performance for competitors
- Developers face unresolvable supply chain risk as they cannot distinguish between model limitations, unsolvable problems, or hidden policy interventions
- Claude Mythos 5 launches simultaneously with safeguards removed, available only to authorized cybersecurity professionals through Project Glasswing