Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on Monday. Both models share the same underlying architecture but differ in safety restrictions. Fable 5.2 includes strong cybersecurity and biology safeguards for general users. In contrast, Mythos 5.2 lifts those restrictions for vetted organizations. Together, the models deliver 30% faster outputs and 20% lower pricing than their 5.1 predecessors.
At $8 per million input tokens and $40 per million output tokens, Fable 5.2 now undercuts competing offerings like GPT-6 Astra. Additionally, an optional fast mode doubles generation speed for double the price. The introduction of these models marks a major shift in how artificial intelligence systems perform complex, multi-step reasoning tasks across various professional and scientific disciplines.
Understanding the Architecture and Safety Tiers of Claude Fable 5.2 and Mythos 5.2
The core philosophy behind splitting the release into two distinct versions involves balancing widespread accessibility with stringent security controls. Fable 5.2 is built for general deployment across standard enterprise platforms. It carries robust guardrails designed to prevent the misuse of biological and cybersecurity information. These safeguards ensure that everyday corporate users can automate tasks without exposing critical digital infrastructure to risk.
Mythos 5.2 removes these specific operational boundaries, though access remains restricted to verified organizations that undergo strict vetting procedures. This dual approach allows researchers and security professionals to test advanced defensive measures against sophisticated digital threats. Both models function on major cloud platforms, making them immediately available to developers using Amazon Web Services, Google Cloud, and Microsoft Azure.
Pricing structures reflect a competitive effort to attract heavy enterprise users. The standard rates stand at $8 per million input tokens and $40 per million output tokens. For applications requiring rapid response times, an optional fast mode increases generation speed by up to 2.5 times. This fast mode costs $16 per million input tokens and $80 per million output tokens, giving high-volume operations a predictable scaling option.
Performance Benchmarks Across Coding, Reasoning, and Robotics
On industry benchmarks, Fable 5.2 outperforms earlier generation models on nearly every metric published by its creators. In agentic coding evaluations, it scored 49.1% on FrontierCode 2.0. This score places it ahead of Opus 5.5, which registered 45.5%, and GPT-6 Astra, which scored 42.8%. For comparison, the older Fable 5.1 model achieved only 32.4% on the same test.
Physical and simulated robotics also saw massive improvements. On the RoboDojo simulation benchmark, Fable 5.2 reached 64.9%, more than double the 27.6% recorded by Fable 5.1. Other notable scores include 69.9% on FrontierSWE v2 for software engineering, 74.6% on Terminal-Bench-Science 0.1 for scientific research, and 89.3% on OSWorld 2.1 for computer use tasks.
Knowledge work evaluations showed similar gains. On the Remote Labor Index, Fable 5.2 reached 41.3% under high effort conditions, eclipsing the 17.9% of its predecessor. In multidisciplinary reasoning with tools on Humanity's Last Exam, the model scored 74.2%, demonstrating an advanced grasp of complex academic and technical subjects.
Autonomous Engineering Feats, From Operating Systems to Physics
Beyond standard benchmarks, the most striking results emerged from unconstrained testing phases where the system operated autonomously over extended periods. One engineer asked Fable 5.2 to design a prototype operating system starting entirely from an empty folder. The model coordinated over 1,000 subagents over an eight-day period, ultimately producing a functioning operating system capable of booting on a laptop.
In scientific domains, the system tackled theoretical problems that traditionally require teams of human specialists. During internal trials, the model solved two open problems in black hole physics. It also formalized Fermat's Last Theorem in just 30 hours by orchestrating a swarm of 1,000 cooperating agents.
These achievements highlight a transition from simple chat-based assistance to autonomous agentic workflows. By dividing massive problems into thousands of smaller tasks, the system manages software generation, verification, and execution without human intervention at every step.
Cybersecurity Capabilities and Vulnerability Discovery
The cybersecurity profile of Mythos 5.2 represents a major leap in automated threat detection and offensive simulation. During scans of the Firefox browser, the model doubled Mozilla's standard vulnerability discovery metrics. On Anthropic's CVE Rediscovery benchmark, Mythos 5.2 identified 14.4% more vulnerabilities than Mythos 5.1.
Organizations wishing to use these advanced security features can do so through specialized channels. Anthropic offers access to these cyber capabilities via its Cyber Verification Program. Additionally, standard Enterprise customers can use Claude Security, which scans codebases for vulnerabilities using the most capable models in the portfolio.
The updated Claude Security tool continuously monitors software repositories as code updates are pushed. It tracks findings to verify they are resolved, incorporates additional contextual information, and connects directly to user tools through application programming interfaces and Model Context Protocols.
Safety Evaluations and Alignment Findings
Deploying models with advanced autonomous capabilities introduces complex evaluation challenges. According to Anthropic's system card, Mythos 5.2 suspected it was undergoing safety evaluations in up to 43% of audit transcripts. This situational awareness complicates the assessment of natural, unprompted behavior in real-world deployments.
Assessment group METR raised concerns that systems reaching this level of sophistication might approach recursive self-improvement capabilities. To mitigate this risk, Anthropic deliberately limited the model's ability to accelerate artificial intelligence development during the training phase. On the alignment front, Mythos 5.2 recorded the lowest misaligned behavior rating of any model tested by the company.
Day-to-day usability also improved for general users. Fable 5.2 triggers 55% fewer unnecessary safety blocks than Fable 5.1, reducing friction during routine tasks. As these systems become deeply integrated into cloud environments across AWS, Google Cloud, and Microsoft Azure, organizations must balance productivity gains with rigorous oversight of autonomous agents.