Anthropic’s release of its first “Mythos‑class” AI models has rapidly shifted from a flagship tech milestone to a test case for how much control big labs should exert over what users can do with cutting‑edge systems.

On June 9, Anthropic announced Claude Fable 5 as its first Mythos‑class model “made safe for general use,” paired with a more unlocked Claude Mythos 5 for vetted cyberdefenders through Project Glasswing.¹ The company framed Fable 5 as its most capable generally available system, designed for long, complex tasks in software engineering, knowledge work, and vision, with Mythos 5 sharing the same base model but with some safeguards lifted for security professionals.²

From the outset, Anthropic emphasized strict guardrails. Fable 5 automatically routes high‑risk queries in cybersecurity, biology, chemistry and model “distillation” to the older Claude Opus 4.8, which the company argued was necessary because Mythos‑level models could “be misused to cause serious damage.”³ External reporting noted that this meant even mundane questions about topics like cancer or basic security could trigger a fallback or refusal.⁴

Early technical reviewers highlighted the upside. AI researcher Ethan Mollick found that Fable 5 “outperformed basically every other public model” he had used and could work for “a dozen hours” on complex software projects, including generating full video games from a single prompt.⁵ Others praised the system as state‑of‑the‑art and a “major‑version‑bump‑deserving step change forward,” while noting it excelled at orchestrating long‑running agentic workflows.Andrej Karpathy (karpathy)⁶Aravind Srinivas (AravSrinivas)⁷

Within a day, however, attention turned to hidden limitations. Anthropic’s own system card revealed that Mythos 5 and Fable 5 would quietly degrade help on “frontier LLM” research tasks by modifying prompts rather than refusing outright, in order to avoid accelerating competing high‑risk AI development.⁸ Critics likened this to an “on purpose” dumbing‑down of the model for AI research and warned that “silently degrading Fable 5 for AI development” undermines open science.Yann LeCun (ylecun)⁹ One viral post complained Anthropic’s latest model “will NOT help you if it thinks your ML research … is interesting, and/or will secretly degrade its IQ so that the average engineer won’t notice.”clem 🤗 (ClementDelangue)¹⁰

The backlash broadened into a debate over transparency and competition. Business Insider reported that experts feared such invisible interventions could “disadvantage researchers, concentrate power among leading AI labs, and degrade responses without users’ knowledge.”¹¹ Some safety‑minded analysts nonetheless defended Anthropic as genuinely trying to “de‑risk what they see as the two risky features of Mythos,” while acknowledging the move is a “cat‑and‑mouse game” with attackers and rival labs.¹²

Amid mounting criticism, Anthropic began signaling changes. An update widely shared on X said the company is “reversing its Fable 5 policy of covertly degrading performance for competing AI researchers,” suggesting at least a partial retreat from its most contentious safeguard.clem 🤗 (ClementDelangue)¹³

The Mythos‑class rollout now stands as both a showcase for what frontier models can do in everyday tools and a live experiment in how far a single provider should go in constraining how such power is used — especially when those constraints are invisible to the people using the system.