Anthropic on Monday launched two new artificial intelligence models, Claude Fable 5 and Claude Mythos 5, which share the same underlying brain but are positioned for very different users. Here is how the two models compare.
Both Fable 5 and Mythos 5 belong to a new tier Anthropic is calling “Mythos-class”, which sits above the company’s existing Opus class. Crucially, the underlying model is the same in both cases. The difference between them lies entirely in the safeguards layered on top.
Anthropic explained the naming choice in a footnote to its announcement. Fable comes from the Latin storymeaning “that which is told”, which the company noted is akin to the Greek mythos. The shared etymology reflects the fact that the two models are, at their core, identical.
Fable 5 is available to the general public from today, accessible through the Claude API and Anthropic’s consumer and enterprise interfaces. Anyone can sign up.
Mythos 5, by contrast, is restricted. It is being deployed initially to existing Project Glasswing partners, including the US government and a small group of cybersecurity and critical infrastructure organisations. Anthropic said it also plans to open a separate trusted access programme for biomedical research, in which biology and chemistry safeguards will be lifted for vetted scientists.
Fable 5 includes new classifiers, which Anthropic describes as separate AI systems that detect potentially sensitive queries. The classifiers cover three areas: cybersecurity, including offensive cyber tasks and “agentic hacking”; biology and chemistry, where queries that could provide uplift for bioweapons research are flagged; and what Anthropic calls “distillation”, referring to attempts to extract a model’s capabilities to train rival AI systems.
When the classifiers trigger, the response is automatically handled by Anthropic’s older Claude Opus 4.8 model instead. The company said its early data shows that more than 95 per cent of Fable 5 sessions involve no fallback at all. For those sessions, it said, Fable 5’s performance is “effectively the same as that of Mythos 5”.
Mythos 5 has cybersecurity safeguards lifted for its current vetted user base, with the biology and chemistry track planned separately.
Both models claim state-of-the-art performance across software engineering, knowledge work, vision and scientific research, according to Anthropic.
Fable 5’s capabilities have been showcased primarily through customer testimonials. Payments company Stripe cited a 50-million-line Ruby codebase migration that the model completed in a day, work that it said would have taken a human team over two months. Cursor’s chief executive Michael Truell placed Fable 5 at the top of its CursorBench evaluation, and GitHub chief product officer Mario Rodriguez called it “a real step forward for the developers GitHub serves”.
Mythos 5’s signature capabilities lie in research domains. Anthropic said the model accelerated parts of its internal protein design process by around ten times, with researchers preferring its molecular biology hypotheses to those of Opus-class models 80 per cent of the time in blinded comparisons. One Mythos hypothesis, a novel mechanism for an E. coli protein, was independently corroborated by a separate lab working on the same problem.
Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens, which Anthropic said is less than half the price it had previously charged for Claude Mythos Preview. Subscription users on Pro, Max, Team and Enterprise plans will get Fable 5 included at no extra cost until June 22, after which continued use will require purchased credits.
The two models illustrate how Anthropic is trying to balance two competing priorities: making frontier AI capabilities widely accessible, while limiting the risks that come with those capabilities. Fable 5 is the public-facing version that ordinary users will interact with. Mythos 5 is the unrestricted version reserved, for now, for partners whose work justifies access to the full set of capabilities.
For most users, the practical experience will be identical, since the safeguards trigger in fewer than one in twenty sessions. For partners working on cybersecurity or, soon, biomedical research, the difference will mean having access to a model that does not pause when the questions get sensitive.
