Claude Fable 5.2 API

Technical Specifications of Claude Fable 5.2

What Is Claude Fable 5.2?

Claude Fable 5.2 is Anthropic's frontier AI model designed for demanding technical reasoning, complex software engineering, scientific research, and long-horizon autonomous workflows. Released on November 9, 2026, it advances the Claude Fable series with stronger agentic coding performance, improved computer-use capabilities, more effective reasoning, and the ability to maintain context across exceptionally long tasks.

Unlike models primarily optimized for short conversational exchanges, Claude Fable 5.2 is designed to work through complicated problems over extended sessions. It can investigate unfamiliar codebases, plan multi-stage engineering projects, coordinate tools and subagents, evaluate intermediate results, and produce completed deliverables such as documents, spreadsheets, and presentations.

Its 1-million-token context window makes it particularly useful for applications involving large repositories, extensive technical documentation, lengthy research materials, and complex multi-step conversations. Its maximum output of 128,000 tokens also supports substantial generated responses and technical artifacts.

Claude Fable 5.2 shares its underlying model capabilities with Claude Mythos 5.2. The distinction is primarily in deployment safeguards: Fable includes stronger restrictions on high-risk cybersecurity and biological applications, making it available to general API customers. Mythos is reserved for organizations approved through Anthropic's verification programs.

Main Features of Claude Fable 5.2

• 1-million-token long-context reasoning: Processes extensive technical documents, large codebases, research collections, and long conversation histories within one context window. This is especially useful for repository-level engineering and research workflows where relevant information is distributed across many files.
• Advanced agentic software engineering: Delivers stronger performance on complex coding and repository maintenance tasks. It can investigate code, design implementation strategies, modify multiple components, and work through iterative engineering processes. Its FrontierCode 2.0 and FrontierSWE v2 results demonstrate improvements over Claude Fable 5.1.
• Adaptive thinking with configurable effort: Adaptive thinking is always enabled, while developers can control reasoning depth through five effort settings: low, medium, high, xhigh, and max. Higher effort can be useful for difficult engineering, research, and planning tasks, while lower settings can reduce unnecessary computation.
• Computer-use and tool orchestration: Supports agentic workflows that interact with computer interfaces and external tools. Developers can use the updated computer toolset and coordinate specialized agents for complicated tasks. The model can also function as an advisor to other Claude models or orchestrate smaller models for simpler subtasks.
• Scientific and multidisciplinary reasoning: Demonstrates stronger results in scientific research, multidisciplinary reasoning, robotics simulation, and knowledge-intensive workflows. Its performance on Terminal-Bench-Science and Humanity's Last Exam illustrates its suitability for technically demanding research assistance.
• Improved visual understanding: Accepts image inputs and can interpret visual information such as screenshots, diagrams, and charts. Its 73.4% Chartography result provides a measurable indication of its chart-recognition capabilities.

Benchmark Performance of Claude Fable 5.2

Anthropic published several evaluation results comparing Claude Fable 5.2 with Claude Fable 5.1 and Claude Opus 5.5.

The results indicate several notable improvements:

1. Software engineering: FrontierCode 2.0 improves from 32.4% for Fable 5.1 to 49.1% for Fable 5.2 at Xhigh effort. FrontierSWE v2 increases from 56.3% to 69.9%.
2. Agentic business workflows: AutomationBench improves to 51.4%, compared with 31.4% for the previous Fable generation.
3. Scientific research: Terminal-Bench-Science 0.1 reaches 74.6%, substantially exceeding the previous generation's 52.6%.
4. Computer interaction: OSWorld 2.1 reaches 89.3%, demonstrating strong performance in computer-use evaluations.
5. Visual interpretation: Chartography reaches 73.4%, compared with 44.8% for Fable 5.1.

These are Anthropic-reported evaluation results, not guarantees of success on individual customer workloads. Benchmark scores also depend on evaluation methodology, effort settings, and tool availability.

Claude Fable 5.2 vs. Other Claude Models

Claude Fable 5.2 belongs to the frontier technical-work category. The most useful comparisons are with Claude Fable 5.1, Claude Opus 5.5, and Claude Sonnet 5.5.

Claude Fable 5.2 vs. Claude Fable 5.1: Fable 5.2 brings substantial gains in coding, computer use, scientific research, and reasoning benchmarks. It also introduces API compatibility changes involving preserved thinking blocks and the computer-use toolset, which developers should review before upgrading.

Claude Fable 5.2 vs. Claude Opus 5.5: Fable 5.2 is the stronger choice when a task requires unusually demanding technical reasoning or long-running agentic execution. Opus 5.5 remains a sensible starting point for most ordinary coding and knowledge-work tasks, according to Anthropic.

Claude Fable 5.2 vs. Claude Sonnet 5.5: Sonnet is more appropriate for faster, general-purpose tasks and workflows where the additional reasoning capacity of Fable is unnecessary. Fable is better suited to complex technical projects requiring extended planning, verification, and orchestration.

Limitations of Claude Fable 5.2

Despite its advanced capabilities, Claude Fable 5.2 has several important limitations:

1. Cybersecurity and biology restrictions: Fable includes safety classifiers that block many high-risk cybersecurity activities, including exploit generation and certain penetration-testing tasks, as well as most high-risk biological research requests. Approved organizations requiring these capabilities should investigate Claude Mythos access.
2. Adaptive thinking cannot be disabled: Developers cannot switch off adaptive thinking. They can control reasoning depth through the effort parameter.
3. Preserved-thinking compatibility: Earlier conversation messages and tool definitions must remain consistent with the context that generated the model's thinking blocks. Modifying prior turns can trigger a 400 error.
4. Computer-use migration requirements: The older computer_20251124 tool is not supported on most platforms. Developers should migrate to computer_toolset_20260801, subject to platform-specific exceptions.
5. Structured tool orchestration constraints: Forced tool selection using tool_choice types any and tool returns a 400 error. Developers should design tool-calling workflows around the supported selection modes.
6. Data-retention requirements: Fable 5.2 requires 30-day data retention and is not available under zero-data-retention arrangements unless Anthropic expressly authorizes an exception.

Best Use Cases for Claude Fable 5.2

1. Large-scale software engineering: Repository-wide refactoring, architectural migrations, debugging, and complex codebase maintenance.
2. Autonomous development agents: Long-running coding agents that plan tasks, invoke tools, inspect intermediate results, and iteratively improve implementations.
3. Technical and scientific research: Multistep investigations, scientific literature analysis, technical hypothesis development, and research synthesis.
4. Business process automation: Complicated workflows involving document analysis, structured information extraction, spreadsheet creation, and multiple external tools.
5. Computer-use automation: Interacting with software interfaces, navigating applications, and performing multistep computer-based tasks.
6. Multimodal technical analysis: Interpreting screenshots, charts, diagrams, and visual technical documentation.