What Claude Mythos 5 and Claude Fable 5 Are
Claude Mythos 5 and Claude Fable 5 are simultaneous frontier AI models from Anthropic that share a common architecture but differ in access controls, aiming to separate general use from tightly governed high-stakes applications while still delivering state-of-the-art performance in reasoning, coding, and specialized professional domains. Anthropic frames Claude Fable 5 as the public-facing version of its ‘Mythos-class’ foundation, while Claude Mythos 5 is the more capable, less restricted variant available only under controlled-access programs. Both models build on months of Mythos Preview testing, where internal evaluations showed unprecedented cybersecurity abilities and long-horizon agentic behavior. This dual-model AI model release signals a move away from a single monolithic flagship towards a portfolio approach, where safety, governance, and capability are intertwined product decisions rather than separate research and policy tracks.
From Oceanus to Mythos 5: Frontier Reasoning and Cybersecurity
The path to Claude Mythos 5 runs through the Oceanus checkpoint, identified as claude-oceanus-v1-p and red teamed in early June. Oceanus extended the Mythos line beyond chat into advanced reasoning capabilities, long-horizon agentic workflows, coding, and cybersecurity, with early testers reporting markedly stronger outputs even before full optimization. Those experiments fed into the production Mythos 5 and Fable 5 systems. On benchmarks, the gains are clear: on SWE-Bench Pro, Anthropic says the models score 80.3%, ahead of Claude Mythos Preview at 77.8% and Opus 4.8 at 69.2%. They reach 29.3% on FrontierCode Diamond, more than double Opus 4.8’s 13.4%. Red teaming around Oceanus helped Anthropic discover not only performance improvements but also risk patterns, especially in autonomous vulnerability discovery, and those lessons appear to have shaped which capabilities are locked behind Mythos 5’s restricted access.
A Two-Tier Access Model: Who Gets Mythos 5 vs Fable 5
Anthropic’s dual-model approach rests on different access levels rather than different core technology. Claude Fable 5 and Claude Mythos 5 share the same underlying foundation and pricing: USD 10 (approx. RM46) per million input tokens and USD 50 (approx. RM230) per million output tokens. The distinction lies in governance. Fable 5 is meant for general Claude users across coding, reasoning, and professional tasks, acting as the public window into the new architecture. Mythos 5 keeps more of the controversial cyber reasoning features that emerged in Mythos Preview, which had identified thousands of software vulnerabilities and showed autonomous vulnerability discovery. As a result, Anthropic is limiting Mythos 5 to selected cybersecurity organizations, infrastructure operators, and strategic partners via Project Glasswing, extending a controlled-access pattern first used when Mythos Preview was restricted to about 200 organizations.
Implications for Claude Users and the Product Ecosystem
For everyday Claude users, Claude Fable 5 will be the default encounter with Anthropic’s latest frontier AI models, offering leading scores across reasoning, coding, tool use, and agent benchmarks such as AutomationBench and OSWorld-Verified. Power users and enterprises will see a clearer gradient of capability: Fable 5 for broad deployment, Mythos 5 for heavily reviewed, higher-risk work. This formalizes a tiered, capability-based distribution model inside the Claude ecosystem, where frontier reasoning and cybersecurity powers are concentrated in a smaller set of vetted customers rather than across all Pro or Max tiers. It also sets expectations that future Mythos-class systems may appear first in restricted form, only later gaining more general exposure once safety evaluations, red teaming, and real-world monitoring reduce uncertainty about misuse and failure modes.
Competitive and Safety Positioning in the Frontier AI Race
Anthropic’s timing places Claude Mythos 5 and Claude Fable 5 squarely in the current frontier AI race, alongside rivals such as GPT-5.5 and Gemini 3.1 Pro, which the company’s benchmarks say the new models surpass in key software, agent, and reasoning tests. This launch follows an Institute paper arguing that AI is already speeding up AI development itself, citing a 52x training-optimization speedup with Mythos Preview, while warning that the field has not yet reached recursive self-improvement and should seek ways to slow development when necessary. The controlled release of Mythos 5 via Project Glasswing lets Anthropic commercialize its most capable model while presenting a safety-first posture. Rather than a single flagship for everyone, the firm is signaling that future AI model releases will likely come as stratified families, balancing market pressure with more deliberate deployment.
