What Claude Fable 5 Is and Why Anthropic Built It
Claude Fable 5 is Anthropic’s publicly released Mythos-class AI model, designed as a safer, more tightly controlled system that trades some direct access to high-risk capabilities for stronger guardrails, stricter logging, and a failover mechanism, with the aim of delivering state-of-the-art performance in coding, reasoning, and knowledge work while reducing cybersecurity and misuse risks. Fable 5 originates from the Mythos line that Anthropic initially held back, warning that an earlier Mythos system could become a powerful hacking aid. The company now describes Fable as separate from Mythos 5 mainly because of its safeguards, not its raw capability. Many requests involving cybersecurity, biology, chemistry, or distillation are routed to the older Claude Opus 4.8, which Anthropic considers safer to expose directly. In effect, Fable 5 is Anthropic’s attempt to ship frontier performance as a safe AI model without repeating the panic that followed the Mythos preview.
Red Teaming Mythos: How Safety Shaped the Model
Anthropic’s work on Mythos-class systems started with extensive red teaming, giving outside testers early access to a checkpoint known as claude-oceanus-v1-p. That preview stage focused on advanced reasoning, software engineering, cybersecurity, and long-horizon agentic tasks rather than casual chat, signaling that Anthropic expected the sharpest safety issues in exactly those domains. According to TestingCatalog, red-team access typically precedes a broader launch by a week or two, and early community experiments suggested a noticeable performance jump over existing models. Internally, Anthropic paired this with research arguing that Mythos Preview delivered a 52x training-optimization speedup, but framed the result as a warning about acceleration rather than a milestone to celebrate. That backdrop explains why Fable 5 arrives wrapped in additional classifiers, fallbacks, and usage controls: the company is reacting to concrete, tested failure modes rather than hypothetical risk alone.
Safety Features in Practice: Guardrails, Classifiers, and Failover
Anthropic positions Claude Fable 5 as “tamer” because it embeds multiple layers of control. The most visible is the automatic failover to Claude Opus 4.8 on prompts that fall into high-risk categories such as cybersecurity, biology, chemistry, or model distillation. Opus is tuned to avoid the more dangerous behaviors Anthropic saw in Mythos testing, so Fable’s raw power is rarely exposed where misuse would be easiest. On top of that, Fable 5 ships with new classifiers—separate AI models trained to detect problematic content and potential misuse before an answer is returned. Anthropic’s model safety card notes that Mythos 5 achieved a 4.8 percent success rate for prompt injection attacks across 100 attempts, comparable to Opus 4.8, suggesting improvements but not invulnerability. In practice, safer AI here means more refusals, more rerouting, and more automated judgement calls about what users are allowed to ask.
Pricing, Data Retention, and the Cost of Safer AI
Anthropic is asking customers to pay more for these Anthropic safety features. Claude Fable 5 and Mythos 5 are priced at USD 10 (approx. RM46) per million input tokens and USD 50 (approx. RM230) per million output tokens, rates that the company says are less than half those for Claude Mythos Preview, which cost USD 25 (approx. RM115) per million input and USD 125 (approx. RM575) per million output. At launch, Fable 5 is temporarily included at no extra cost for Pro, Max, Team, and seat-based Enterprise plans until June 22, after which it will shift to usage credits if capacity allows. Alongside the Claude Fable 5 release, Anthropic also tightened its data retention policy: even organizations that previously had zero data retention now face a 30-day log retention period for prompts and outputs on Mythos-class models, used for trust and safety but not for training.
IPO Optics and the Meaning of ‘Safer’ for Anthropic
The timing of Claude Fable 5’s debut matters as much as its technical profile. Anthropic is preparing for an IPO at a reported valuation of USD 1 trillion (approx. RM4.6 trillion), and needs to convince both investors and regulators that it can ship increasingly capable systems without opening the door to new classes of harm. The Mythos saga—first withheld as too dangerous, then reintroduced as a constrained, safer AI model—offers a narrative that Anthropic can govern its own frontier work. But the tradeoffs are real. The New York Times notes that while hackers may find it harder to abuse Fable, defenders may also lose access to the most capable cybersecurity tooling. For power users, safer AI now means higher token costs, stricter logging, and more indirect access to Mythos-level capability, all justified as the price of keeping the model’s most dangerous behaviors in check.
