Claude Fable 5.2 and Mythos 5.2: Features and Pricing

Anthropic has announced Claude Fable 5.2 and Claude Mythos 5.2, two new Artificial Intelligence models designed to tackle complex technical tasks across software development, scientific research, engineering and cybersecurity. The company says the models offer faster output generation, stronger benchmark performance and improved capabilities for handling challenging problems that require extended reasoning and limited human intervention.

The launch highlights Anthropic’s continued focus on developing AI systems that can do more than answer questions or generate basic code. The company is positioning its latest models for demanding professional applications, including complex software migrations, scientific investigations, robotics simulations and vulnerability detection.

According to Anthropic, Claude Fable 5.2 and Mythos 5.2 are built on the same underlying model architecture but operate with different safety protections and access restrictions. Fable 5.2 is intended for broader use, while Mythos 5.2 is designed for vetted organisations working on sensitive cybersecurity and biomedical research projects.

Anthropic also claims that both models can generate tokens up to 30% faster than their respective previous versions. Alongside the performance improvements, the company has introduced standard and optional fast-mode API pricing, giving developers different options depending on their speed and cost requirements.

Claude Fable 5.2 and Mythos 5.2: Key Highlights

• Anthropic says the new models deliver up to 30% faster token generation than Claude Fable 5.1 and Mythos 5.1.
• Both models target advanced coding, reasoning, scientific research, engineering and robotics tasks.
• Fable 5.2 achieved 49.1% on the FrontierCode 2.0 benchmark at Xhigh effort, according to Anthropic.
• Mythos 5.2 identified 89.6% of vulnerabilities in Anthropic’s CVE Rediscovery evaluation.
• Standard API pricing starts at $8 per million input tokens and $40 per million output tokens.

Claude Fable 5.2 and Mythos 5.2: What’s New?

The latest models are designed to improve performance in tasks that require multiple steps, technical knowledge and the ability to work through complicated problems. Anthropic says the improvements cover agentic coding, computer use, robotics and scientific research.

Agentic coding allows an AI system to work through software development tasks with greater independence. Instead of simply suggesting code snippets, an agent can plan changes, interact with development tools, identify errors and work towards completing a larger programming objective.

Anthropic’s reported testing results suggest that Fable 5.2 is intended to handle more demanding engineering work. The company says the model helped engineers complete complex software migrations and design three working electronic devices. These examples point towards applications in which AI systems can contribute to technical workflows involving several connected stages.

The models have also demonstrated progress in scientific research. Anthropic reports that Fable 5.2 made advances on theoretical physics problems that had previously remained unresolved. However, the announcement does not provide enough detail to independently assess the scientific significance of those results or the extent of human involvement.

Anthropic Claims 1,000 AI Agents Formalised Fermat’s Last Theorem Proof in 30 Hours

One of the most notable claims in Anthropic’s announcement concerns the use of multiple AI agents to formalise a proof of Fermat’s Last Theorem. According to the company, a team of 1,000 Fable 5.2 agents completed the formalisation process in approximately 30 hours, with minimal human intervention.

Fermat’s Last Theorem is a famous mathematical problem stating that no three positive integers can satisfy the equation aⁿ + bⁿ = cⁿ for any integer value of n greater than 2. Its proof, established by mathematician Andrew Wiles in the 1990s, is one of the major achievements in modern mathematics.

Formalising a mathematical proof involves translating mathematical reasoning into a precise language that a proof assistant or computer system can verify. This process requires more than producing a convincing explanation: each logical step must satisfy the rules of the formal system being used.

Anthropic says an earlier attempt using an unreleased Claude model took 11 days, compared with the 30-hour result reported for Fable 5.2. The comparison suggests a substantial improvement in the company’s multi-agent mathematical workflow.

However, the result should be understood in the context of formal proof verification rather than as a claim that the AI independently discovered the original proof of the theorem. The announcement also does not provide enough information about the verification environment, the division of work among agents or the precise level of human involvement to draw broader conclusions about autonomous mathematical research.

Claude Fable 5.2 Coding Performance Compared With Rival AI Models

Coding performance is one of the major areas in which Anthropic is highlighting improvements. The company says Fable 5.2 achieved a score of 49.1% on the FrontierCode 2.0 benchmark at Xhigh effort.

Anthropic’s reported comparison places the model ahead of Opus 5.5, which scored 45.5%, and GPT-6 Astra, which recorded 42.8% on the same benchmark.

The results indicate that Fable 5.2 performed better than the two listed competitors in this particular evaluation. However, benchmark scores do not necessarily represent how an AI model will perform across every programming language, software environment or real-world development project.

Actual coding productivity also depends on factors such as code quality, debugging reliability, the ability to understand existing systems and the amount of human review required before changes can be deployed. Developers evaluating these models will need to consider those practical factors alongside benchmark performance.

Anthropic’s focus on complex software migrations is particularly relevant for businesses maintaining large applications. Such projects can involve changes to dependencies, infrastructure, programming frameworks and interconnected components. An AI model capable of understanding these relationships could help engineering teams reduce repetitive work and identify potential problems earlier in the development process.

Robotics Performance: Fable 5.2 Scores 64.9% on RoboDojo Sim

Anthropic is also highlighting improvements in robotics-related tasks. According to the company, Claude Fable 5.2 achieved a 64.9% success rate on the RoboDojo Sim robotics benchmark.

Robotics applications require AI systems to interpret information, reason about physical environments and select actions that can achieve specific objectives. Unlike conventional text-based tasks, robotics workflows often involve constraints related to movement, object interaction and the physical properties of the environment.

Performance in a robotics simulation can help evaluate whether a model is capable of supporting tasks involving these challenges. However, success in a simulated environment does not automatically establish equivalent performance with physical robots operating in unpredictable real-world conditions.

The reported RoboDojo Sim result nevertheless adds to Anthropic’s broader effort to position its models for applications beyond conventional chatbots and software development. Better reasoning and computer-use capabilities could make advanced AI systems more useful in engineering environments where digital planning and physical operations intersect.

Mythos 5.2 Focuses on Cybersecurity and Vulnerability Detection

Cybersecurity is a central part of Anthropic’s positioning for Claude Mythos 5.2. The company says the model has improved capabilities for identifying and reproducing software vulnerabilities, with a score of 89.6% in its CVE Rediscovery evaluation.

Vulnerability detection involves identifying weaknesses in software that could potentially be exploited by attackers. Reproducing a vulnerability can help security researchers understand how a flaw works, assess its potential impact and develop appropriate fixes.

These capabilities can be useful for defensive security teams responsible for reviewing applications, testing systems and protecting digital infrastructure. AI-assisted analysis may help researchers examine large amounts of code and investigate potential weaknesses more efficiently.

At the same time, advanced vulnerability research carries security risks. The same technical capabilities that help defenders identify software weaknesses could potentially be misused to develop or improve cyberattacks. This is one reason Anthropic is restricting access to Mythos 5.2 rather than making it generally available.

According to the company, Mythos 5.2 is intended for organisations that meet its eligibility requirements through dedicated verification programmes. The additional restrictions reflect the sensitive nature of the cybersecurity and biomedical research applications for which the model is designed.

Why Anthropic Is Offering Two Versions of the Same Underlying Model

Although Claude Fable 5.2 and Mythos 5.2 share the same underlying model, Anthropic says they have different safeguards. The distinction is based on how the models are intended to be used and the risks associated with their capabilities.

Claude Fable 5.2 includes additional protections for cybersecurity and biology-related tasks and is available to general users through Anthropic’s platforms. These safeguards are intended to support broader access while limiting potentially dangerous applications.

Claude Mythos 5.2, in contrast, is designed for vetted organisations working on sensitive cybersecurity and biomedical research. Its restricted distribution gives Anthropic a way to make advanced technical capabilities available in controlled environments.

This approach reflects a broader challenge facing AI developers: increasingly capable models can provide significant benefits in areas such as software security and scientific research, but they can also create risks when used without appropriate oversight.

By separating general availability from restricted access, Anthropic is attempting to balance the usefulness of its latest models with the need for additional controls in sensitive technical fields.

Claude Fable 5.2 and Mythos 5.2 API Pricing Explained

Alongside the performance improvements, Anthropic has announced API pricing for the new models. The standard pricing starts at $8 per million input tokens and $40 per million output tokens.

Input tokens represent the information sent to the model, such as user instructions, source code, documents and conversation history. Output tokens refer to the content generated in response, including explanations, code and other results.

For applications that require faster responses, Anthropic is offering an optional fast mode. The company says this mode can increase token generation speeds by up to 2.5 times compared with the standard mode.

Fast mode comes at a higher price, with API rates of $16 per million input tokens and $80 per million output tokens.

The pricing structure gives developers a choice between standard processing and a faster option. Applications that can tolerate longer response times may prefer standard pricing, while time-sensitive workflows could benefit from the additional speed offered by fast mode.

Businesses should also consider how frequently their applications send large prompts and generate lengthy responses. Since input and output tokens are priced separately, total API costs will depend on usage patterns rather than the headline price alone.

Claude Fable 5.2 Availability Across Anthropic’s Platforms

Claude Fable 5.2 is available through Anthropic’s platforms and major cloud infrastructure providers, including Amazon Web Services, Google Cloud and Microsoft Azure.

Developers can also access the model through the Claude Platform using the model identifier claude-fable-5-2. This gives development teams a way to integrate the model into applications and technical workflows through the API.

Its availability across multiple platforms could make it easier for organisations already using different cloud providers to evaluate and adopt the model. Developers can explore its coding, reasoning and computer-use capabilities without necessarily changing their existing infrastructure arrangements.

Claude Mythos 5.2 Availability: Verification Required

Claude Mythos 5.2 has a more restricted availability model. Anthropic says eligible organisations can access it through the Cyber Verification Program and Life Sciences Verification Program.

These programmes are intended for organisations working in areas where advanced AI capabilities require additional oversight. Cybersecurity researchers may use the model to investigate vulnerabilities and strengthen software security, while eligible life sciences organisations can access its capabilities for biomedical research.

Enterprise customers can also access Mythos-powered security capabilities through Claude Security. This provides a route for businesses to use Anthropic’s security-focused AI capabilities in their defensive security workflows.

Unlike Fable 5.2, Mythos 5.2 is not positioned as a general-purpose model for unrestricted public access. Organisations interested in using it must meet Anthropic’s eligibility and verification requirements.

What the Launch Means for AI-Powered Software Development and Research

The launch of Claude Fable 5.2 and Mythos 5.2 highlights the growing competition among AI developers to build systems capable of completing complex technical work. The focus is shifting beyond conversational responses towards models that can operate across longer workflows, use tools and contribute to specialised professional tasks.

For software developers, improvements in agentic coding could make AI assistants more useful for maintaining large codebases, migrating applications and investigating difficult technical problems. If models can reliably complete more stages of a development task, engineering teams may be able to spend less time on repetitive implementation work and more time reviewing architecture and product decisions.

Scientific research is another area where these capabilities could have an impact. The reported mathematical formalisation work illustrates how multiple AI agents may be coordinated to handle complex tasks. Similar approaches could support researchers in organising calculations, checking logical steps and exploring technical problems, although meaningful scientific conclusions will still require rigorous verification.

Cybersecurity presents both opportunities and challenges. Faster vulnerability discovery could help organisations identify weaknesses before attackers exploit them. However, powerful systems that can reproduce vulnerabilities also require safeguards, access controls and responsible deployment practices.

Anthropic’s decision to distinguish between Fable and Mythos demonstrates how model availability may increasingly depend on the sensitivity of the capabilities involved. As AI systems become more capable, developers may need to consider not only performance and price but also who can access particular tools and under what conditions.

Claude Fable 5.2 vs Mythos 5.2: Main Differences

What to Expect From Anthropic’s Latest AI Models

Claude Fable 5.2 and Mythos 5.2 represent Anthropic’s effort to make AI systems more capable in technical environments where accuracy, reasoning and the ability to complete multi-step tasks matter. The company’s reported results cover several areas, from coding and robotics to mathematical formalisation and cybersecurity research.

The claimed improvements in generation speed and the introduction of fast-mode pricing also give developers additional options when building applications around the models. Meanwhile, the different access policies for Fable and Mythos show that Anthropic is treating sensitive technical capabilities differently from general-purpose AI features.

However, benchmark results and company demonstrations are only part of the picture. Real-world reliability, independent evaluations, operating costs and the amount of human oversight required will determine how useful these models become for businesses and researchers.

For now, Fable 5.2 offers the broader route for developers and general users to explore Anthropic’s latest capabilities, while Mythos 5.2 is aimed at organisations working in more sensitive research and security environments. Their adoption will depend on whether the reported improvements translate into dependable results across practical applications.