claude / fable-5

Overview

Claude Fable 5 is a high-capability Claude model from Anthropic. It uses adaptive thinking by default, then returns a final text answer you can act on. It supports very large prompts, so it can keep long documents, specs, and chat history coherent. This helps when you need fewer handoffs between “analyze” and “deliver.”

What you can build

• Long-horizon coding help for multi-file refactors and repository-level changes
• Document Q&A on PDFs, including charts and tables
• Extraction of structured facts from reports, contracts, and policies
• Screenshot and diagram interpretation for debugging and UI reviews
• Research workflows that cite sources when web search is enabled
• Multi-turn assistants that keep context using a stable user and session identifier

Inputs

• A required instruction you write as plain text. This is the question or task you want done.
• Optional attachments, limited to images (JPEG, PNG, GIF, WebP) and PDF documents. You can attach multiple files per request (up to 50 on this listing).
• If you attach an animated GIF, the model uses only the first frame.
• Images can be large, but platforms often enforce practical limits. Claude’s vision docs describe a maximum image dimension of 8000×8000 pixels, plus per-platform file-size limits.
• PDFs can be processed as visual documents. Claude’s PDF support docs describe request-level limits such as total request size and maximum pages per request. Dense PDFs can hit the context window before page limits.
• An optional system-level instruction that sets global behavior for the session. Use it for tone, formatting rules, and “always/never” constraints.
• Optional user and session identifiers (string or number). Provide these if you want the app to store and reload chat history.
• An optional effort level that controls how much work the model does before answering. Higher effort increases depth on hard tasks.
• An optional web search toggle. Enable it when the task needs current facts beyond the model’s training cutoff.
• An optional randomness control for generation. Note that Claude Fable 5 may reject non-default sampling settings, depending on the underlying API mode.
• A maximum output budget (up to 128k tokens). This caps the total response length.

Outputs

• A text response that typically uses Markdown for structure, lists, and code blocks.
• If you enable web search, the response can include citations that point to the sources used.
• For PDFs and images, the response can describe visual elements, summarize content, and extract key fields into readable tables.

Recommended settings

• Use high effort for most complex work. Step down to medium or low for routine Q&A.
• Use xhigh effort for the hardest tasks, like multi-stage plans and large refactors.
• Set a large output budget for high-effort tasks. Thinking and the final answer share that budget.
• Keep image and document attachments to 20 or fewer per request when possible. Many-image requests can trigger stricter per-image limits on some platforms.

Limitations

• Claude Fable 5 has safety classifiers that can refuse certain requests. Refusals can be more frequent than on older Claude models.
• Non-default sampling settings such as temperature changes can be incompatible with Fable 5’s thinking mode in some API configurations.
• Images above the model’s native vision resolution are downscaled before analysis. Very small text can become unreadable after downscaling.
• Animated images aren’t fully supported. Only a single frame is analyzed.
• PDF understanding depends on vision-style processing. Scanned PDFs, skewed pages, or heavy compression can reduce accuracy.
• Low-quality or non-structured inputs can cause missing fields, misread tables, and incorrect extractions. This is common with blurry screenshots, dense spreadsheets rendered as images, and multi-column scans.

Safety & compliance

• The model is designed to block high-risk dual-use requests, especially in areas like cybersecurity and biology.
• Plan for refusals in user experience. Treat them as valid outcomes and offer safer alternatives.
• Don’t use outputs as professional advice without expert review. This matters most for legal, medical, and financial decisions.
• Data handling rules can vary by platform. Anthropic designates Claude Fable 5 as a covered model with specific retention requirements on supported services.