Fable 5: The Full Story from Capabilities to Drama (Ep. 1002 with Jon Krohn)
Anthropic's highly anticipated Claude Fable 5 model, a public version of its advanced "Mythos class" AI with state-of-the-art capabilities in software, vision, and long-context tasks, was released and then swiftly pulled offline by the U.S. government after just three days. The removal, initiated as an export control action over national security concerns stemming from a disputed "jailbreak" claim, highlights the growing tension between frontier AI development, AI safety, and regulatory oversight.
This episode delves into the release and subsequent rapid withdrawal of Anthropic’s Claude Fable 5 model, described as the most capable AI model ever released to the public. The narrative weaves together two main threads: what made Fable 5 exceptional and why it was taken offline by the U.S. government after only three days.
Claude Fable 5: A New Tier of AI Capability
Anthropic had previously structured its models into tiers: Haiku (small, fast, cheap), Sonnet (capable middle), and Opus (powerful, expensive top). Earlier this year, Anthropic introduced a new “Mythos class” tier above Opus. The first model in this class, Claude Mythos Preview, was deemed too dangerous for general release and was only made available to a select group of cyber defenders and critical infrastructure providers through “Project Glasswing” due to its potential for exploiting software vulnerabilities. Fable 5 represents Anthropic’s attempt to provide a public version of this advanced technology.
Fable 5 and its restricted sibling, Mythos 5, are based on the same underlying model, differing only in their safeguards. The nomenclature is intentional: ‘Fable’ derives from the Latin fabula (“that which is told”), mirroring the Greek mythos, indicating a shared root with different guardrails. Mythos 5, with lifted cyber safeguards, was given to Glasswing partners, while Fable 5, with firmly switched-on safeguards, was released to the general public.
Unprecedented Performance and Capabilities
Fable 5 demonstrated state-of-the-art performance across nearly all AI benchmarks, showing particular strength in software engineering, knowledge work, vision, and scientific research. A key differentiator is its ability to maintain coherence and perform across long, multi-step tasks – the longer and more complex the task, the greater Fable’s lead over previous models.
Concrete examples highlight its prowess:
- Software Engineering: Stripe utilized Fable 5 to perform a codebase-wide migration on a 50-million-line Ruby codebase in a single day, a task estimated to take a full team of engineers over two months manually.
- Vision: Unlike earlier Claude models that needed elaborate helper tools, Fable 5 beat the video game Pokemon Fire Red using only raw screenshots and a minimal harness.
- Memory/Long Context: In the deck-building game Slay the Spire, with access to persistent file-based notes, Fable 5 boosted its performance three times more than the Opus model. This suggests the model is adept at “getting better at taking notes for itself and acting on them later.”
Pricing and Safety Classifiers
Reflecting its advanced capabilities, Fable 5 was priced at $10 per million input tokens and $50 per million output tokens, roughly twice the cost of Claude Opus 4.8, yet less than half the price of the original Mythos Preview. This pricing aligns with the broader trend of decreasing cost-per-capability in AI.
Fable 5 implemented stringent safeguards via separate AI systems called classifiers. These classifiers monitor queries in sensitive domains: cyber security, biology/chemistry, and “distillation” (attempts to extract model capabilities for competitive training). If a request is flagged, Fable 5 quietly defaults to Opus 4.8, and the user is informed. Anthropic designed these classifiers to be “aggressively conservative,” acknowledging they “err on the side of caution” and sometimes flag harmless requests. While the safeguards trigger in fewer than 5% of sessions, false positives have generated “grumbling” among early users, a concern Anthropic aims to address.
These safeguards are a direct response to the dual-use nature of Mythos-class capabilities, where beneficial applications (e.g., hardening systems, gene therapy) could be maliciously exploited (e.g., breaking into systems, developing bioweapons). Anthropic’s strategy was to provide broad access to benefits while containing risks, even conducting an external bug bounty program that reportedly found no universal way around the safeguards.
The Government Intervention and Aftermath
Days before Fable 5’s release, Anthropic publicly advocated for a “coordinated break pedal” for frontier AI development, citing risks of recursive self-improvement. This juxtaposition of warning about powerful AI and then releasing its most powerful public model underscored the company’s internal tension: ship capability, but carefully.
However, this cautious approach did not prevent intervention. On a Friday evening, Anthropic received a directive from the U.S. federal government, ordering an immediate worldwide shutdown of both Fable 5 and Mythos 5 due to national security concerns. The models were pulled within hours, lasting only three days.
Export Control and Disputed Jailbreak
The directive was an export control action issued by Commerce Secretary Howard Lutnik. It designated Fable 5 and Mythos 5 as subject to export controls, prohibiting access by any foreign national, whether abroad or within the U.S. Due to the broad interpretation covering even foreign-born Anthropic employees, the company concluded that the only practical way to comply was to shut down access globally.
The government’s underlying concern, as understood by Anthropic, was a claimed jailbreak of Fable 5 reportedly demonstrated by Amazon and communicated by its CEO Andy Jassy to the White House. Anthropic, however, strongly disputes the severity of this finding. They described the technique as a “narrow, non-universal jailbreak” that merely involved asking the model to read a codebase and identify flaws. Anthropic argued that the vulnerabilities found were previously known, minor, discoverable by other public models (like OpenAI’s GPT-5.5), and routinely found by legitimate security professionals. Essentially, Anthropic contended that the restricted capability was neither unique to Fable 5 nor the “doomsday scenario” it was being treated as.
Broader Implications and Ongoing Dispute
This incident highlights a complex interplay between frontier AI development, national security, and commercial pressures. The irony that Anthropic’s own warnings about powerful AI may have inadvertently contributed to the government’s swift action was not lost. This isn’t the first clash; the Department of Defense had previously labeled Anthropic a “supply chain risk,” and reports suggested earlier attempts to stop Fable’s launch.
The timing was particularly sensitive for Anthropic, which had confidentially filed for an IPO, reportedly targeting a valuation near a trillion dollars. Having its flagship model pulled by its own government during this period presents an undeniable business challenge.
While Anthropic complied, it publicly expressed its belief that the government’s decision was incorrect. Legal proceedings are underway, with some agencies reportedly allowed to continue using Anthropic’s tools. The saga is far from over. There’s even a speculative upside for Anthropic’s brand, as the government’s ban could inadvertently signal that Anthropic is at the forefront of AI development, with rivals playing catch-up.
For users, Fable 5 and Mythos 5 remain unavailable, with queries defaulting to Opus 4.8. Their return, and under what conditions, hinges on the resolution of this ongoing dispute between Anthropic and the U.S. government.