Anthropic launches Fable 5.2 and Mythos 5.2 with specialized access

New models for specific tasks

Anthropic recently introduced two new models, Fable 5.2 and Mythos 5.2, to its lineup. These versions share the same underlying architecture but are designed for different types of users and safety requirements. Fable 5.2 is available to the general public, while Mythos 5.2 is restricted to vetted organizations.

The release comes five months after the initial launch of the Fable 5 and Mythos 5 models. While Fable 5.2 is open to all users, it includes specific safeguards to address biological and cybersecurity risks. In contrast, Mythos 5.2 is reserved for entities working in life sciences and cyber defense.

Anthropic is positioning these models as specialists rather than universal replacements. The company suggests that Fable 5.2 is best suited for the most difficult open-ended technical tasks. For most other types of knowledge work and coding, the company still recommends using its Opus 5.5 model.

Performance and pricing logic

The pricing for Fable 5.2 is set at $8 per million input tokens and $40 per million output tokens. Users who need faster results can opt for a fast mode that doubles these prices to $16 and $80 per million tokens. This faster mode generates tokens up to 2.5 times quicker than the standard version.

Anthropic argues that the primary value of Fable 5.2 is not the price per token, but the cost per completed task. The company claims that Fable 5.2 roughly matches the performance of Opus 5.5 on the FrontierCode 2.0 benchmark. This performance is achieved at approximately half the cost per task compared to the larger model.

On the FrontierSWE v2 coding benchmark, Fable 5.2 received a score of 69.9 per cent. This score is higher than the 62.3 per cent recorded for Opus 5.5 and the 65.5 per cent reported for OpenAI’s GPT-6 Astra. These figures are based on internal reports from Anthropic and have not been independently verified.

Security and restricted access

Mythos 5.2 is specifically tailored for cybersecurity teams and security researchers. During internal testing of Common Vulnerabilities and Exposures (CVE), the model found 10.2 per cent more known security flaws than Opus 5.5. Because of these capabilities, the model is not available to the general public.

The restricted distribution of Mythos 5.2 reflects the risks associated with models that can identify software flaws. Claude Enterprise customers can access this specific model through a dedicated security portal. Reports indicate that India’s national cybersecurity agency, CERT-In, has already received access to the model.

Anthropic also improved its safety classifiers with this release. The company reports that the new filters are roughly 55 per cent more accurate at allowing legitimate user requests compared to the previous Fable 5.1 version. This update aims to reduce friction for users while still maintaining necessary safety boundaries for high-risk tasks.

Enterprise data governance

To address the concerns of large companies, Anthropic introduced Enterprise Frontier Safeguards. This feature is designed for organizations that have very strict requirements regarding how their data is handled. It allows for safety monitoring to occur within the customer's own infrastructure rather than on Anthropic's servers.

This setup enables zero-data-retention arrangements, which means the AI provider does not keep any of the customer's information. This is a significant move for enterprise buyers who want to use AI without risking the privacy of their internal data. The service is currently in a private beta phase.

Several major companies, including Salesforce, Stripe, and Bridgewater, are currently participating in this private beta. Their feedback will help Anthropic determine if this infrastructure-based monitoring meets the needs of large-scale corporate environments. The goal is to balance safety monitoring with strict data governance.

The multi-model strategy

The launch of these models suggests a shift toward a multi-model strategy for businesses. Instead of using one expensive model for every single request, organizations may begin to tier their AI usage. This involves using different models for routine work, complex technical problems, and high-security tasks.

Technology leaders must weigh the cost of tokens against the cost of completed tasks. A more expensive model might actually be cheaper if it requires fewer attempts or less human intervention to get the correct result. However, these savings will vary depending on the specific workload and the infrastructure costs involved.

Organizations looking to adopt these new tools should test them against their own specific use cases. Because the performance data is based on Anthropic's own claims, independent testing is necessary to see how the models behave in real-world scenarios. The goal is to find the right balance between performance, cost, and security for every unique task.