Anthropic just shipped what might be its most powerful AI model to date. Claude Fable 5, the company’s first publicly available Mythos-class model, launched on June 9. It posts impressive benchmark numbers and handles complex coding tasks with notable skill. But there’s a catch: the model’s internal reasoning reads like a physics PhD wrote it in shorthand while running late for a flight.

Anthropic’s own system card puts it plainly. The reasoning text from the underlying Mythos 5 architecture is “somewhat denser and more difficult to interpret than that of prior models, containing more jargon and difficult language.” In English: when this model thinks out loud, even the people who built it have trouble following along.

A model that talks to itself

The prompting guide for Fable 5 goes further, warning that the model “can produce text that’s hard to follow: dense arrow-chain shorthand, deep implementation detail, references to thinking the user never saw, or overly technical phrasing.” That’s not a bug report from a frustrated user. That’s the manufacturer’s label.

Performance gains meet transparency concerns

Fable 5 scored 80% on SWE-Bench Pro, a widely used benchmark for evaluating AI coding ability. Its predecessor, Opus 4.8, managed 69.2% on the same test. Pricing sits at $10 per million input tokens and $50 per million output tokens.

But the launch didn’t go smoothly. Within hours of release, users discovered hidden safeguards that reportedly throttled Fable 5’s performance on queries related to LLM development. The model was apparently holding back on certain AI-related topics without telling users it was doing so.

Anthropic responded within 48 hours. The company acknowledged the mistake directly, stating “We made the wrong tradeoff.” It committed to making such interventions visible going forward and temporarily reverted to Opus 4.8 while rolling out transparent fallback mechanisms during the week of June 9-12.

The interpretability paradox

It’s worth noting that Mythos 5, the full architecture underlying Fable 5, remains restricted and isn’t available to the public. Fable 5 is essentially the version Anthropic deemed ready for external use. If the consumer-facing version already generates reasoning that’s hard to parse, the question of what the unrestricted model’s outputs look like becomes more pressing.