Anthropic launches Claude Fable 5.2 for complex technical tasks on AWS

A new model for complex work

Anthropic has officially released Claude Fable 5.2 and Claude Mythos 5.2 on Amazon Bedrock and the Claude Platform on AWS. These models are specifically designed to handle high-risk and technically demanding workloads that require more than simple conversation. Fable 5.2 is positioned as the company's most capable model for tasks involving coding, mathematics, and the sciences.

Unlike standard models that might struggle with ambiguity, Fable 5.2 is built to interpret complex prompts and iterate on them. It can run unattended for periods ranging from hours to several days to complete a single objective. The model is designed to catch its own errors and optimize its approach as it progresses through a task.

Anthropic notes that Fable 5.2 is expected to behave similarly to their internal alignment testing results. In those tests, the model showed improved reliability and fewer hallucinations compared to previous versions like Opus 5.5 and Sonnet 5.5. It is intended to surface important information more consistently than its predecessors.

Specialized use cases and domains

Claude Fable 5.2 is intended for a variety of high-level technical applications. This includes agentic coding across large, multi-repository codebases and coordinating many subagents for long-running projects. It is also suited for scientific research, including formal mathematical proofs and new analytical methods.

The model can also be applied to financial analysis for quantitative research and forecasting. Additionally, it supports design work for 3D environments, hardware, and motion simulations generated through code. These capabilities make it a versatile tool for engineering and research teams that require deep reasoning.

In contrast, Claude Mythos 5.2 is restricted to two specific high-risk domains. It is available only through Anthropic’s verification programs for cyber security and life sciences. These use cases include security engineering, red teaming, and penetration testing, as well as biomolecular design and research synthesis.

Safety and data privacy

Anthropic has designated Claude Fable 5.2 as a Covered Model, which carries specific safety requirements. By default, using this model requires data retention for up to 30 days and a human safety review. On Amazon Bedrock, Amazon personnel perform these reviews within the AWS boundary to ensure security.

To manage these requirements, users can set the aws_review data retention mode on their account or project before invoking the model. For organizations with stricter privacy needs, Anthropic and AWS collaborated on Enterprise Frontier Safeguards (EFS). This system combines the privacy of zero data retention (ZDR) with advanced misuse detection.

Currently, eligible customers can use Fable 5.2 with zero data retention until June 30, 2027. A future version of EFS that keeps data in a user-controlled bucket with their own keys is expected to launch later. This ensures that high-risk technical work can be performed without compromising sensitive data.

Developer tools and migration

Developers can access Fable 5.2 through the Amazon Bedrock console or via the Anthropic Messages API. Anthropic also provided a migration tool called the claude-api skill for teams already using Claude models on Bedrock. This open-source skill helps users upgrade their existing codebases to the new model efficiently.

The migration skill automatically handles several technical updates, such as swapping model IDs and adjusting thinking settings. It can also clean up deprecated parameters and suggest appropriate effort levels based on the specific use case. It detects various deployment platforms to ensure the correct model ID formats are used during the transition.

Users can also interact with the tool conversationally to improve cache hit rates or add context compaction to their agents. These features help reduce costs on Amazon Bedrock by optimizing how the model handles long-running sessions. The skill is bundled in several popular tools, including JetBrains IDEs and Warp, in addition to being available in Claude Code.

Technical nuances and billing

One specific technical feature for developers involves how the system handles declined queries in high-risk domains. If a cyber security query is declined by Fable 5.2, the user can retry it on a different model like Opus 4.8. However, the user's code must manage this retry logic rather than the system switching models automatically.

To prevent unnecessary costs, Anthropic introduced a fallback credit token for requests that hit the prompt cache. If a request is refused, the refusal includes this token, which can be sent with the retry. This ensures the retry is billed as a cache read rather than a new write, saving the user from paying twice for the same data.

The models are available today on Amazon Bedrock through specific inference profiles in various regions. Pricing for these models is managed through the standard Amazon Bedrock pricing structure. Developers can begin testing the model immediately using the provided Getting Started notebook on GitHub.