Claude Fable 5 Complete Tutorial: Get Started with Loop Engineering
Claude Fable 5 is Anthropic’s flagship model released on June 9, 2026. It targets a 1M context, autonomous multi-day operation without supervision, and xhigh-effort tier reasoning. Fable 5 was temporarily paused on June 12 due to export controls by the Ministry of Commerce, then the ban was lifted and it went back online on June 30. It is currently Anthropic’s highest-tier model available via API. This article summarizes the key points from Anthropic’s official “Prompting Claude Fable 5” handbook, breaking down the core behavior changes in Fable 5 compared with Opus 4.8, and how to adjust the corresponding prompts and harnesses. It is a complete first Chinese guide for getting started with Fable 5, suitable for developers, AI agent builders, and enterprise technology decision-makers.
What is Fable 5: 1M context, xhigh effort, multi-day autonomous operation
Fable 5 API model: claude-fable-5. It also has the following specifications:
- Context window: 1M tokens (million-level context). Maximum output per call: 128K tokens.
- Pricing: Input $10/M tokens, Output $50/M tokens (prompt caching can get a 10% discount).
- Local reasoning in the United States: 1.1× price surcharge.
- Available platforms: Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, Microsoft Foundry.
Adaptive thinking:
- Unique “thinking” mode; cannot be disabled.
- thinking.display can be set to summarized (read the summary) or omitted (do not read, default).
Safety classifier:
- Fable 5 includes a built-in refusal classifier. Requests involving attack-oriented cybersecurity, biological and life sciences, or requests to extract model reasoning will be refused (HTTP 200 but stop_reason: refusal).
- Mythos 5 has no such classifier; it is only available to approved Project Glasswing customers.
Data retention: 30 days (not available for ZDR zero data retention).
Anthropic’s official positioning of Fable 5 improvements over Opus 4.8 is in seven areas:
1) Long-horizon autonomy
2) Higher first-shot correctness for complex problems
3) Vision
4) Enterprise workflows
5) Code review and debugging
6) Handling ambiguous requirements
7) Concurrent sub-agents
The shared underlying logic behind these seven improvements is one thing: Fable 5 can continuously and stably produce outputs during the time when the user isn’t watching.
Core paradigm shift: From Prompt Engineering to Loop Engineering
The biggest mindset change with Fable 5 isn’t writing better prompts—it’s changing from prompt steps to designing loops. With Opus 4.6/4.7, the mindset was: “Write a good prompt; Claude returns a good answer.” With Fable 5, the mindset is: “Design an environmental loop that lets Claude self-correct, repeatedly cycling through plan → act → review → improve until the task is complete.”
In the official handbook, Anthropic explicitly provides this loop design instruction:
“Establish a method for checking your own work at an interval of [X] as you build. Run this every [X] interval, verifying your work with subagents against the specification.”
A translation of this passage:
“Establish a self-check mechanism to validate your own work at intervals of [X]. Every [X] interval, run verification and check your work against the specification using subagents.”
The key is to use a set of independent context verification subagents (fresh-context verifier subagents), not let the main agent critique itself—because the main agent already has bias.
This shift also responds to the previously reported “Harness Engineering” concept: the next battleground for AI is not the model, but the layer of architecture outside the model; and Akshay’s viewpoint that “the model is just a node in the loop.” Loop engineering is Anthropic’s official version of this line of thinking.
Effort parameter grading:
5-level control of how many tokens Fable uses
Effort is Fable 5’s most important cost-control knob. It is divided into 5 levels, from highest to lowest, affecting how many tokens Claude is willing to spend on thinking, tool calls, and text responses:
- max: highest capability, no cost ceiling. Used for frontier problems requiring the deepest reasoning. Frequent use will generate large costs, not recommended for routine use.
- xhigh: expanded capabilities, suitable for durations over 30 minutes, token usage...”
Claude Fable 5 is Anthropic’s flagship model released on June 9, 2026. It targets a 1M context, autonomous multi-day operation without supervision, and xhigh-effort tier reasoning. Fable 5 was temporarily paused on June 12 due to export controls by the Ministry of Commerce, then the ban was lifted and it went back online on June 30. It is currently Anthropic’s highest-tier model available via API. This article summarizes the key points from Anthropic’s official “Prompting Claude Fable 5” handbook, breaking down the core behavior changes in Fable 5 compared with Opus 4.8, and how to adjust the corresponding prompts and harnesses. It is a complete first Chinese guide for getting started with Fable 5, suitable for developers, AI agent builders, and enterprise technology decision-makers.
What is Fable 5: 1M context, xhigh effort, multi-day autonomous operation
Fable 5 API model: claude-fable-5. It also has the following specifications:
- Context window: 1M tokens (million-level context). Maximum output per call: 128K tokens.
- Pricing: Input $10/M tokens, Output $50/M tokens (prompt caching can get a 10% discount).
- Local reasoning in the United States: 1.1× price surcharge.
- Available platforms: Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud, Microsoft Foundry.
Adaptive thinking:
- Unique “thinking” mode; cannot be disabled.
- thinking.display can be set to summarized (read the summary) or omitted (do not read, default).
Safety classifier:
- Fable 5 includes a built-in refusal classifier. Requests involving attack-oriented cybersecurity, biological and life sciences, or requests to extract model reasoning will be refused (HTTP 200 but stop_reason: refusal).
- Mythos 5 has no such classifier; it is only available to approved Project Glasswing customers.
Data retention: 30 days (not available for ZDR zero data retention).
Anthropic’s official positioning of Fable 5 improvements over Opus 4.8 is in seven areas:
1) Long-horizon autonomy
2) Higher first-shot correctness for complex problems
3) Vision
4) Enterprise workflows
5) Code review and debugging
6) Handling ambiguous requirements
7) Concurrent sub-agents
The shared underlying logic behind these seven improvements is one thing: Fable 5 can continuously and stably produce outputs during the time when the user isn’t watching.
Core paradigm shift: From Prompt Engineering to Loop Engineering
The biggest mindset change with Fable 5 isn’t writing better prompts—it’s changing from prompt steps to designing loops. With Opus 4.6/4.7, the mindset was: “Write a good prompt; Claude returns a good answer.” With Fable 5, the mindset is: “Design an environmental loop that lets Claude self-correct, repeatedly cycling through plan → act → review → improve until the task is complete.”
In the official handbook, Anthropic explicitly provides this loop design instruction:
“Establish a method for checking your own work at an interval of [X] as you build. Run this every [X] interval, verifying your work with subagents against the specification.”
A translation of this passage:
“Establish a self-check mechanism to validate your own work at intervals of [X]. Every [X] interval, run verification and check your work against the specification using subagents.”
The key is to use a set of independent context verification subagents (fresh-context verifier subagents), not let the main agent critique itself—because the main agent already has bias.
This shift also responds to the previously reported “Harness Engineering” concept: the next battleground for AI is not the model, but the layer of architecture outside the model; and Akshay’s viewpoint that “the model is just a node in the loop.” Loop engineering is Anthropic’s official version of this line of thinking.
Effort parameter grading:
5-level control of how many tokens Fable uses
Effort is Fable 5’s most important cost-control knob. It is divided into 5 levels, from highest to lowest, affecting how many tokens Claude is willing to spend on thinking, tool calls, and text responses:
- max: highest capability, no cost ceiling. Used for frontier problems requiring the deepest reasoning. Frequent use will generate large costs, not recommended for routine use.
- xhigh: expanded capabilities, suitable for durations over 30 minutes, token usage...”