Anthropic Drops Claude Fable 5, Crypto Twitter Immediately Declares Mutiny 🪖

Anthropic on Tuesday launched Claude Mythos 5, a restricted-access AI model with cybersecurity capabilities, alongside Claude Fable 5, a safeguarded version designed for the general public. The company said Mythos 5 will initially be available only to approved cybersecurity organizations, critical infrastructure operators, government partners and selected life sciences researchers, with broader access to follow "in consultation with the U.S. government." The release expands the Mythos program that began in April under Project Glasswing, when Anthropic released a preview to a limited group of cyber defenders and critical software infrastructure providers. The U.K.'s AI Security Institute earlier reported that a preview version became the first AI model to complete a 32-step corporate network intrusion exercise without human assistance, and Mozilla reported in April that a preview version of Mythos discovered over 271 vulnerabilities in the Firefox browser. Anthropic said last month that Mythos uncovered more than 10,000 high or critical-severity vulnerabilities in "systemically important software," including around 6,200 high or critical-severity vulnerabilities in more than 1,000 open-source projects it investigated.

Claude Fable 5 uses the same core design as Mythos 5 with added safeguards, and Anthropic said it was "made safe for general use," with guardrails that reroute certain topics, including cybersecurity, to a different model, Claude Opus 4.8. The company added that "releasing a model this capable comes with risks. Without safeguards, Fable 5's capabilities in areas like cybersecurity could be misused to cause serious damage." The launch comes as Anthropic has confidentially filed for an IPO with the SEC at a reported potential valuation of up to $1 trillion, with rival OpenAI also said to have confidentially filed. The release also follows President Donald Trump's AI executive order proposing that AI developers provide the government with access to models they consider "covered frontier models," and reports that Anthropic has embedded engineers at the National Security Agency to help deploy Mythos for cybersecurity operations reportedly targeting China and Iran.

Crypto users reacted with alarm. Simon Dedic, founder of venture firm Moonrock Capital, posted on X that with Fable 5, the "cost and skill required to find exploitable flaws in smart contracts is about to drop to basically zero," and recommended users revoke wallet approvals, remove value from protocols and move crypto to fresh hardware wallets. Curve Finance co-founder Michael Egorov pushed back, arguing that smart contracts contain only a few thousand lines of code compared with the millions in the software where Mythos found bugs, and that "both humans and 'usual' AI perfectly fit that code in context and can reason well about it." He said he suspected "we might not be having a wave of DeFi code hacks, but we may see a lot of things in OpSec getting hacked (looking like multisig keys compromises) and supply chain attacks on frontend dependencies." In April, the value of crypto stolen in hacks hit $629.7 million, the highest since February 2025, which analysts linked to the use of AI.

Fable 5 launched to sharp criticism over its cost and token usage. The model is priced at $10 per million input tokens and $50 per million output tokens, double the cost of Claude Opus 4.8, and counts double against subscription usage limits. Bleeping Computer reported that Fable drained a $100 Max subscription's daily allowance in just under nine minutes, while Scrimba CEO Per Borgen posted on X that "just tried Fable. It burned 1.3M tokens in 7 minutes. That's $160 per hour. Equivalent to a $333k/year salary." Theo from T3 Chat said he had spent over $1,000 in tokens in one day on his $200 subscription plan, and Pixelated Ink CTO Josh Ellithorpe said Fable 5 "burns tokens like no other model," limiting him to a few prompts before draining his quota. Anthropic said Workflow mode, the heaviest feature, breaks complex prompts into parallel subagent tasks that cost more compute by design, and that a new system prompt of around 120,000 tokens is loaded into every new conversation.