anthropic Claude Platform Release Notes ·

Anthropic Launches Claude Fable 5 and Mythos 5 with 1M Token Context

aipreviewengineer
feature announcement

Anthropic has released Claude Fable 5, their most capable widely available model, and Claude Mythos 5 for Project Glasswing participants. Both models feature a 1 million token context window and enhanced adaptive thinking capabilities. These updates impact users interacting with the Claude API, particularly concerning tokenizer changes, safety classifier responses, and data retention policies for Fable 5.

  • Claude Fable 5 and Claude Mythos 5 Models Released
  • Enhanced Safety Classifiers with Refusal Handling
  • Scheduled Deployments for Claude Managed Agents
  • Environment Variable Credentials for Managed Agents Vaults
  • New Tokenizer Increases Token Count for Same Text
Features (4)
  • Claude Fable 5 and Claude Mythos 5 Models Released

    Claude Fable 5 is now widely available, and Claude Mythos 5 is available for Project Glasswing participants. Both models support a 1 million token context window, 128k max output tokens, and always-on adaptive thinking.

  • Enhanced Safety Classifiers with Refusal Handling

    Claude Fable 5 runs safety classifiers on requests and responses, returning 'refusal' stop_reason for declined requests, with no billing for pre-output refusals. An opt-in 'fallbacks' parameter is in beta to re-run refused requests on other models.

  • Scheduled Deployments for Claude Managed Agents

    Claude Managed Agents now support scheduled deployments, allowing sessions to run on a cron schedule without requiring users to manage their own scheduler.

  • Environment Variable Credentials for Managed Agents Vaults

    Claude Managed Agents vaults now support environment variable credentials, enabling secure injection of secrets into the agent's sandbox for various CLI and SDK authentications.

Enhancements (4)
  • New Tokenizer Increases Token Count for Same Text

    Claude Fable 5 and Mythos 5 utilize a tokenizer similar to Claude Opus 4.7, resulting in approximately 30% more tokens for the same text compared to older models. A token counting API is available to measure prompts under the new tokenizer.

  • New Refusal Reason for Reverse Engineering

    The 'stop_details.category' field now includes 'reasoning_extraction' for Claude Fable 5, indicating requests blocked due to restrictions on reverse engineering or duplicating model outputs.

  • Adaptive Thinking is Default and Only Mode

    On Claude Fable 5 and Mythos 5, adaptive thinking is the default and only supported mode. Manual extended thinking budgets and assistant prefill are not supported.

  • Configurable Thinking Output Display

    The 'thinking.display' field defaults to 'omitted' on Claude Fable 5 and Mythos 5, similar to recent Opus models. Users can set it to 'summarized' for readable thinking summaries.

Read the original announcement →

https://platform.claude.com/docs/en/release-notes/overview