Claude Fable 5.2 Text Reasoning Model

Overview

Claude Fable 5.2 is a Claude family model made by Anthropic. It’s designed for demanding reasoning and long-horizon agent work. You can give it large amounts of text and images, and it produces text outputs like plans, code, and written deliverables. It’s useful when smaller models lose the thread across long, multi-step tasks.

It supports a 1,000,000-token context window and up to 128,000 output tokens per response. The model uses always-on adaptive thinking, so you control depth with an effort setting.

What you can build

• Multi-step codebase migrations with a long planning and execution loop
• Deep code reviews that track decisions across many files and iterations
• Research assistants that read PDFs, extract key claims, and draft reports
• Vision-assisted analysis like chart reading, UI screenshot critique, or document figure review
• Long-form writing like design docs, RFCs, and project proposals with consistent terminology
• Agentic task plans with checklists, risks, and verification steps

Inputs

• A required instruction written in plain text that describes the task and constraints.
• Optionally, a chat-style message history with roles and message content for multi-turn work. Use this instead of a single instruction when you want a transcript-style conversation.
• Optional attachments (up to 50 files total): JPEG, PNG, GIF, and WebP images; PDF documents; plain-text files; and Markdown files. Use these to provide source material the model should read.
• Optional system-level instructions written in plain text to set global behavior, tone, and boundaries for the run.
• Optional identifiers (a user ID and a session ID, as a string or a number) to keep chat history tied to a specific user and session.
• An optional effort level that controls how much reasoning the model applies before answering. You can choose from low, medium, high, xhigh, and max.
• An optional switch to let the model use web search for up-to-date information when your platform enables it.
• An optional output length limit that caps the response size, up to 128,000 tokens.
• Optional tool-call scaffolding for continuing prior runs, where you pass a prior task token and the matching tool results so the model can continue from the stored turn.

Outputs

The model returns a text response. It commonly includes structured Markdown with headings, lists, tables, and code blocks when appropriate.

For long tasks, the response often contains a clear plan plus detailed work product, such as: - refactored code snippets, - step-by-step procedures, - test plans, - risk and dependency lists, - drafts of documents you can paste into a wiki or PR description.

Recommended settings

• Use High effort for most demanding reasoning and long-horizon agent work.
• Move up to XHigh only for the hardest tasks where you measured a gain on your own evals.
• Use Medium or Low where quality stays acceptable and you want shorter outputs.
• Reserve Max for capability ceiling tests and the rare tasks where token use is not a concern.

Limitations

• Claude Fable 5.2 is a text-output model. It does not generate images, audio, or video.
• Adaptive thinking is always on. You can’t fully disable it, only reduce depth with effort.
• Some platforms reject forced tool use. Expect “auto” or “none” style tool choice only.
• This model’s reliable knowledge cutoff is June 2026. It may miss events after that date unless web search is enabled.
• Low-quality inputs reduce accuracy. Scanned PDFs, blurry screenshots, and noisy images can cause missing or incorrect details.
• Very long tasks can hit your output limit. When that happens, ask for a continuation or a smaller slice.

Safety & compliance

Claude Fable 5.2 includes stronger safeguards in cybersecurity and biology domains. The model may refuse or limit responses for requests that meaningfully enable harm.

Don’t use it to create or improve malware, exploit chains, or instructions for wrongdoing. Don’t request guidance for developing, acquiring, or deploying biological threats. For sensitive use cases, follow your organization’s review and approval process.