Claude Fable 5 und Mythos 5 Release Image © AnthropicClaude Fable 5 und Mythos 5 Release (Image © Anthropic)

Technical Specifications and Pricing Structure

Both models feature a context window of 1 million tokens and a maximum output of 128,000 tokens. Adaptive reasoning is integrated as a standard feature, allowing users to adjust the depth of reasoning via an effort parameter. Pricing remains consistent with that of Claude Fable 5, with one significant exception**: the cost of cache reads has been reduced. Cache reads now cost 0.025 times the base input price—a significant reduction compared to the 0.1 multiplier applied by other models. This change specifically targets long-running agentic sessions that frequently rely on cached data.

Expanded Feature Set

The 5.1 updates focus on six main areas of improvement:

  • Agent-based programming: Improved performance when refactoring multiple files, debugging, and migrating long-running sessions.
  • Knowledge management: Improved ability to create documents, spreadsheets, and presentations from scratch.
  • In-Depth Research: Greater accuracy in multi-step web searches and iterative exploration tasks.
  • Visual Analysis: Better processing of complex tables, documents, and charts in PDF files, supported by cropping and zooming features.
  • Context Analysis: Improved ability to draw conclusions and make connections between data points across the entire 1 million-token window.
  • Computer Usage: More reliable operation of desktop applications and browsers, including improved error handling.

Implementation of New Features

Several new features have been introduced, some of which are currently in beta:

Adjusting Effort Level During Conversations

Users can now adjust the effort level during a session without invalidating the prompt cache. This enables complex reasoning for challenging tasks and low-effort processing for routine steps.

System Messages Limited to a Single Round of Conversation

A new beta feature allows system messages to be limited to a single round of conversation. By setting the clear_at parameter to next_user_message, instructions apply only to that round of conversation and are subsequently removed without altering the conversation history or affecting the prompt cache.

Updates on the Progress of Tool Calls

To improve transparency during long agent rounds, a new “display**: updates” option allows users to receive text-based status updates between tool calls, while the primary inference chain remains hidden.

Content Origin

All text generated by these models contains a statistical watermark. In addition, image and video files generated via the code execution tool contain signed C2PA content certificates when accessed via the Files API. ## Operational and Fundamental Changes Developers migrating from Claude Fable 5 should be aware of three critical fundamental changes:

  1. Tool Selection Restrictions**: Forced tool usage is no longer supported. Requests where tool_choice is set to any or specific tool names will result in a 400 error. Users must use the auto setting and provide explicit instructions in the prompt to trigger tool calls.
  2. Thinking Block Compatibility**: Compatibility is one-way. While Fable 5.1 can read “Thinking Blocks” from earlier models, earlier models cannot read blocks generated by Fable 5.1.
  3. History Integrity**: Editing, reordering, or removing rounds preceding a “Thinking Block” now invalidates that block. To maintain validity, conversations should be treated as “append-only.”

Behavioral Adjustments

The 5.1 models exhibit different behavioral patterns compared to their predecessors:

  • Tool Calls: Parallel tool calls are more variable; the model may now execute individual calls where it previously bundled multiple ones.
  • Information Retrieval: When the effort is low, the model is more likely to access internal memory rather than trigger search or retrieval tools.
  • Output Style: The text is generally denser, with longer sentences and fewer paragraph breaks. Additionally, a significant decrease in the use of bold text, headings, and lists can be observed in chat interfaces.
  • File editing: The model is more likely to rewrite entire files when making small changes, rather than performing targeted edits.

Claude Fable 5.1 and Mythos 5.1 Availability and Deployment

Claude Fable 5.1 is available via the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Access to Claude Mythos 5.1 requires approval from the account teams at Anthropic, AWS, or Google Cloud. A 30-day data retention policy applies to both models.