"The advantage will belong to the side that can get the most out of these tools," Anthropic warned in April when it announced the Mythos model.
Mythos lineage: power, risk, and a cautious rollout
Anthropic’s new release, Claude Fable 5, is drawn from the same underlying model class as Mythos — a model the company previously described as state-of-the-art and potentially dangerous if misused. When Anthropic introduced Mythos, it warned the model was powerful enough to help bad actors attack public and private software, and that in the short term attackers could benefit if frontier labs were not careful about how they released these systems.
Fable 5’s guardrails and the Opus 4.8 fallback
Anthropic says it has built strong guardrails into Fable 5 that block or divert sensitive queries — for example, questions involving offensive cybersecurity, biology, or chemistry — to an earlier model, Opus 4.8. In Anthropic’s framing, Fable 5 is a safer, restricted-release variant of the same model class; by contrast, Claude Mythos 5 is described as the unrestricted version, with those guardrails lifted. Anthropic also previously limited access to Mythos-class models to cybersecurity experts and trusted companies because of the security risks.
Availability, pricing window, and compute cost
Anthropic is offering Fable 5 free for a limited time and says the model is expensive to run because it requires substantial compute. Specifically, Anthropic told customers that until June 22 Fable 5 will be available to all Pro, Max, and Enterprise customers; after that date the company plans to switch to usage-based pricing. Anthropic also explained that the company cannot afford to make Fable 5 as widely available as Opus 4.8 or its earlier models, due to the model’s compute cost.
Token use, Workflow mode, and real-world testing
BleepingComputer tested Fable 5 and reported that the model consumes tokens faster than any other Anthropic model. The publication observed that Fable 5 uses "a massive amount of tokens in a span of minutes," and that running the new Workflow execution system — which breaks complex prompts into smaller tasks and can spin up parallel subagents — can accelerate that consumption sharply. In BleepingComputer’s account, using Workflow with the model set to high thinking exhausted the author’s $100 Max subscription daily usage in nine minutes. The outlet also reported that even in casual interaction, Fable 5 consumes tokens about twice as fast as the Opus model.
Claude Mythos 5 access and trusted partners
Anthropic has kept the unrestricted Claude Mythos 5 available only to a highly vetted group of trusted partners because of the potential for misuse. The company cited government cyberdefenders and specific life sciences researchers as examples of the kinds of organizations granted access to that unrestricted build. That selective access follows Anthropic’s earlier move to limit Mythos-class models to cybersecurity experts and trusted companies in order to reduce the risk that attackers would use the technology to find and exploit vulnerabilities — the company explicitly noted such tools could be abused to discover defects in applications like Firefox.
What this means for government cyberdefenders, Pro/Max/Enterprise customers, and life sciences researchers
- Government cyberdefenders: Anthropic’s approach offers a path to privileged access for teams that need powerful, unrestricted capability, while keeping the general release guarded. Those defenders will continue to be among the candidate audiences for Claude Mythos 5 and will likely need to negotiate the vetted-access process Anthropic has described.
- Pro, Max, and Enterprise customers: Customers in those tiers have a narrow window — until June 22 — to try Fable 5 without incurring usage charges, but should be wary that the model can deplete subscription quotas rapidly, especially when used with Workflow at higher effort settings.
- Life sciences researchers: Anthropic lists "specific life sciences researchers" among potential trusted partners for unrestricted Mythos access, implying a continued split between guided, guardrailed public builds and narrow, vetted research access for sensitive domains.
Anthropic’s statement that it may continue to change capacity — "the company is known for nerfing its models and increasing the capacity later," as BleepingComputer put it — leaves room for adjustments to access and pricing in the coming weeks. For now, organizations and researchers must balance the benefit of cutting-edge model capabilities against rapid token consumption, compute cost, and the company’s deliberate limits intended to reduce misuse.
Original story: https://www.bleepingcomputer.com/news/artificial-intelligence/anthropic-rolls-out-claude-fable-5-but-its-available-for-a-limited-time/
