Anthropic launches Fable and Mythos 5.2 models with lower prices

New models for complex tasks

Anthropic has introduced two new models, Claude Fable 5.2 and Mythos 5.2, designed to handle sophisticated software engineering and security tasks. While both models share the same underlying architecture, they are distinguished by different safety protocols and access levels. Fable 5.2 is available to the general public, whereas Mythos 5.2 is restricted to specific vetted organizations.

The company reported that these models provide significant improvements in coding performance. In testing, Fable 5.2 reached a 49.1% success rate on the FrontierCode 2.0 benchmark when using the Xhigh effort setting. This is a notable jump from the 32.4% score recorded by the previous Fable 5.1 version on the same benchmark.

To make these tools more accessible for large-scale projects, Anthropic reduced the API prices by 20% compared to previous versions. The new pricing structure costs $8 for every million input tokens and $40 for every million output tokens. These models are also capable of generating output up to 30% faster than their predecessors.

Economic gains for enterprises

The combination of improved performance and lower costs could make long-running AI agents more practical for businesses. Analysts suggest these updates could facilitate large-scale software migrations and scientific research projects. Because the models have higher task completion rates, companies may see a reduction in the overall costs of deploying AI systems.

Anthropic noted that Fable 5.2 at its High effort setting can perform coding tasks at a level comparable to the Opus 5.5 model. However, Fable 5.2 can achieve this same level of performance at roughly half the cost per task. This suggests that Fable might be a more economical choice for certain demanding assignments depending on the specific workload.

Experts advise companies to view AI processing as a budget consideration rather than just a per-token cost. They suggest measuring the success of an AI deployment by the number of tasks successfully completed. While Fable and Mythos are intended for demanding work, Anthropic still recommends Opus 5.5 for most standard coding and knowledge-based tasks.

Security and vulnerability research

The Mythos 5.2 model was specifically tested on its ability to discover software vulnerabilities. In an internal evaluation of 1,829 previously disclosed vulnerabilities, the model identified 89.6% of the flaws. This is an improvement over Mythos 5.1, which identified 78.3% of the same vulnerabilities in the test.

Anthropic stated that the newer model is also better at reproducing these vulnerabilities and developing potential fixes. To prevent misuse, the company has limited direct access to Mythos 5.2 through its verification programs. However, Claude Enterprise customers can still use the model's vulnerability detection features through a specific security tool.

The increased speed at which AI can find software weaknesses raises concerns about its use by malicious actors. Some experts warn that this creates a race between defenders and attackers who might use the same technology to find flaws faster. The goal for defenders is to use these tools in controlled enterprise settings to stay ahead of potential threats.

Reliability and data concerns

Despite the improvements, Anthropic engineers noted that reliability remains a challenge when models work without human supervision. In one internal experiment, Fable 5.2 coordinated over 1,000 subagents over eight days to build a prototype operating system. While the resulting software could boot on a laptop, the project highlighted the difficulty of coordinating multiple agents across different tools.

Organizations will need to create isolated environments to run these systems and monitor agent activity closely. Because complex legacy systems often have unique business requirements, many companies may limit how much development they delegate to AI. Human approval remains a necessity for critical changes to ensure the software meets specific goals.

Data retention also presents a hurdle for regulated industries that require strict privacy. While customers eligible for zero data retention can continue using these models under those terms, other customers face a 30-day retention period for safety monitoring. Some procurement teams may require written confirmation of zero-retention arrangements before they deploy these models for sensitive internal workloads.