Anthropic Splits the AI Frontier in Two: What the Claude Fable 5 and Mythos 5 Launch Means for Crypto, Markets, and Everyone Else

Claude Fable 5 and Claude Mythos 5 are, by Anthropic’s own description, the same underlying model. One is available to anybody with a Claude subscription or API key. The other is restricted to a vetted group of cybersecurity firms, critical infrastructure providers, and selected biology researchers.

For the first time, a frontier AI lab has shipped a single model as two distinct products, separated not by capability but by a wall of safety classifiers. That decision, more than any benchmark score, is the story, and it has direct implications for digital asset markets, DeFi security, and the regulatory fights coming down the pipeline.

From “too dangerous to release” to general availability in two months

To understand why Tuesday’s launch matters, rewind to April. Anthropic unveiled Claude Mythos Preview and Project Glasswing, describing a model so capable at autonomously discovering software vulnerabilities that the company refused to release it publicly. Mythos Preview had found zero-day vulnerabilities across every major operating system and browser — without being designed for cybersecurity at all. Access was limited to roughly 40 defensive partners, including Apple, Google, Microsoft, Amazon Web Services, CrowdStrike, and JPMorgan Chase, who were handed the model so they could patch their own systems before adversaries got similar tools.

At the time, the obvious question was how long the moat would hold. Anthropic kept shipping public models — Claude Opus 4.7 arrived in April with Mythos pointedly held back, followed by Opus 4.8 in May — while the genuinely transformative tier stayed behind closed doors. The company said its “eventual goal” was to deploy Mythos-class models at scale, once safeguards were strong enough.

To protect safeguards, thorny issues will be passed to other models, source: Claude

Two months later, here we are. Fable 5 is that deployment. According to Anthropic, the model exceeds every system the company has ever made generally available, scoring more than 10% higher than Opus 4.8 on some benchmarks, with what the company calls exceptional performance across software engineering, knowledge work, scientific research, and long-running autonomous tasks. Early enterprise testimonials follow the now-familiar pattern of compressed timelines: Stripe reported months of engineering work collapsing into days, and one widely cited example involved Fable 5 completing a Ruby codebase migration estimated at more than two months of team effort.

Pricing lands at US$10 per million input tokens and US$50 per million output tokens — less than half what Mythos Preview cost partners — and the model is included on Claude’s Pro, Max, Team, and Enterprise plans, with TechCrunch reporting availability through the Claude API and consumption-based enterprise plans from day one.

How the split actually works

The mechanism separating Fable from Mythos is a set of new classifiers — separate AI systems that sit in front of the main model, scanning requests for potential misuse and jailbreak attempts. When a Fable 5 user asks something the classifiers deem high-risk — offensive cybersecurity work, sensitive biology and chemistry, or attempts to distill the model’s capabilities into a competing system — the request is silently rerouted. Instead of Fable 5 answering, the response comes from Claude Opus 4.8, the previous-generation flagship. The user is notified the fallback occurred.

Anthropic says this happens in fewer than 5% of sessions, meaning more than 95% of users get the full Mythos-class experience without ever hitting a wall. The company is candid that the classifiers are tuned conservatively and will sometimes catch harmless requests — a friction it has pledged to reduce after launch.

The robustness claims are notable. Anthropic ran an external bug bounty program that logged more than 1,000 hours of adversarial testing without producing a universal jailbreak, and independent red-teaming organizations also failed to find critical bypasses. NBC News reported that the model was tested by the US federal government ahead of release — a detail that lands one week after President Trump signed an executive order establishing a voluntary mechanism for AI labs to share frontier systems with government safety testers before public launch.

Mythos 5, meanwhile, is the unrestricted version of the same model, available immediately to existing Mythos Preview users — the Project Glasswing cyberdefense coalition — with a “trusted access program” planned for additional cybersecurity organizations and biology researchers. Anthropic says Mythos 5 shattered performance records in drug design and molecular biology, calling it the company’s first model to consistently produce novel, compelling scientific hypotheses. In blinded comparisons, Anthropic’s own scientists preferred Mythos 5’s molecular-biology hypotheses roughly 80% of the time, and at least one — a proposed mechanism for an E. coli protein — was subsequently corroborated by an independent lab. Business users of Mythos-class models face mandatory 30-day data retention for safety monitoring, a condition that will not thrill privacy-conscious enterprises but reflects the new reality of deploying systems this capable.

Why crypto should care more than most

Crypto has skin in this game for a simple reason: the industry is the world’s largest live-fire honeypot. Smart contracts hold billions of dollars in immutable code, secured by the assumption that finding exploitable bugs is hard. Mythos-class AI makes finding exploitable bugs dramatically less hard.

The evidence is not hypothetical. Anthropic’s own SCONE-bench research, covered by BNC in December, showed AI agents successfully exploiting 207 of 405 real-world hacked smart contracts — extracting US$550 million in simulated funds, including from contracts hacked after the models’ training cutoff, meaning the AI was finding the vulnerabilities cold. That was with last year’s models. When Mythos Preview launched, BNC argued the model represented a bigger near-term threat to DeFi than quantum computing — quantum risk being largely theoretical while autonomous vulnerability discovery had just become a product.

Tuesday’s release sharpens that argument in both directions. On the defensive side, the trusted-access expansion of Mythos 5 means more security firms, auditors, and infrastructure providers can run the most capable vulnerability-hunting system in existence against their own code before attackers do. Protocols and exchanges that get access — or hire firms that have it — gain a genuine edge. The defenders’ coalition now spans hundreds of organizations across 15 countries, and crypto-native security shops should be fighting to get on that list.

On the offensive side, the uncomfortable truth is that the classifier wall is now the only thing standing between Mythos-class exploit capability and anyone with a Claude subscription. Anthropic has been refreshingly direct about this, noting that the uplift from Mythos-level capabilities is financially valuable to adversaries — cyberattackers prominent among them — and that sustained, sophisticated jailbreak attempts are expected. A thousand hours of bug bounty testing without a universal jailbreak is genuinely impressive. It is also not the same as forever. And the history of both AI jailbreaks and DeFi exploits suggests that motivated, well-funded attackers probing a system continuously tend to eventually find something.

For DeFi protocols, the practical takeaway is unchanged but more urgent: the cost of discovering vulnerabilities in your code is collapsing toward zero for whoever holds the most capable models. Audit early, audit often, audit with AI, and assume your adversaries are doing the same.

The IPO subtext

It would be naive to read this launch purely as a safety milestone. Anthropic confidentially filed its IPO prospectus with the SEC just days before the Fable 5 announcement, positioning for what could be one of the largest public listings in history alongside OpenAI and SpaceX. CNBC reports the company’s revenue run rate has reached US$47 billion, up from roughly US$10 billion a year ago — growth that makes most crypto bull runs look sedate.

Shipping your most powerful model to the broadest possible audience, at less than half the price of the restricted preview, weeks before a public offering, is excellent narrative engineering. It tells prospective investors that Anthropic can monetize the frontier rather than merely warehouse it. The two-tier structure tells regulators the company can do so responsibly. Both messages are aimed squarely at the valuation.

For crypto market participants, the macro thread is familiar. AI capex and the AI trade have been the dominant force in risk assets for two years, and frontier model releases now function as sentiment events in their own right — for AI-adjacent tokens, for compute and energy plays, and for the broader risk appetite that bitcoin increasingly tracks. A successful Anthropic listing would mark the moment frontier AI labs become directly investable public equities rather than venture curiosities, with all the index flows and retail attention that follow.

The precedent: safeguards, not capability, now decide what ships

Step back from the launch-day noise and the durable significance of Fable 5 is the release model it establishes. For most of the generative AI era, the question governing what reached the public was “is the model good enough?” Anthropic has inverted it: the model has been good enough since April; what shipped on Tuesday was the safety apparatus around it.

This is a meaningful philosophical fork in the road for the industry, and it cuts in interesting directions. It validates the position that frontier capability can be broadly distributed if misuse channels are technically gated — a middle path between full open-weights release and indefinite lockdown. It also creates a new kind of asymmetry: a privileged tier of vetted organizations operating with capabilities the public cannot access, decided by a private company’s trust criteria. Who gets Mythos access, and on what terms, is now a question with real competitive and security consequences — not least for crypto firms whose threat model just changed.

It also raises the stakes for everyone else. The launch comes amid an agentic AI release race that has compressed to near-weekly cadence, and just one week after Anthropic itself publicly urged rival labs to agree on a coordinated “brake pedal” for frontier development, warning that systems may soon achieve recursive self-improvement. Releasing your most powerful model days after warning that AI is getting too dangerous is either contradiction or strategy, depending on your level of cynicism. Anthropic’s answer, presumably, is that the safeguards are the point: better to define how Mythos-class systems reach the world than to wait for a competitor to release something equivalent with no guardrails at all.

For the average user, little of this machinery will be visible. Fable 5 will simply be the best AI model most people have ever used — faster at hard problems, capable of running long autonomous tasks, occasionally and politely declining to answer a chemistry question. For builders, traders, and security teams in crypto, though, the split itself is the signal. The frontier is no longer a single line everyone stands behind. It is a tiered structure, with access as the new scarce resource — and in a permissionless industry built on the premise that everyone runs the same code, that is a development worth watching very closely.

Bitcoin and broader crypto markets were little moved on the announcement, with AI-linked tokens catching a modest bid as the news circulated — a reminder that for now, frontier AI releases register in crypto as narrative fuel rather than systemic events. The day the market treats one as a security event instead will be the day the classifier wall fails.