Summary of Key Features
Anthropic has released the Claude Fable 5.1 model, along with a guide to prompts to help developers adapt to the new model's behavior changes. The main highlights include:
- Effort Level Control: A feature to balance cost and performance. There are five levels of effort: low (most cost-effective), medium, high (default), xhigh, and max (most intelligent but also most expensive).
- Cost Savings: The low-level setting of Fable5.1 offers similar results to the high-level setting of Fable5, but at a fraction of the cost. The price for caching and reading has been reduced by 75% (from $1 per thousand words to $0.25), making long conversations more cost-effective.
- Flexible Adjustment: You can temporarily switch to the xhigh level when encountering difficulties and then switch back to high level after completion, without affecting the cache.
1. Effort Level Control: A Smart Way to Save Money
The “effort level” is the core cost-control mechanism in Fable5.1. It's like adjusting the model's “level of effort”:
- Five Levels Available: low (most economical), medium, high (default), xhigh, and max (most intelligent but most expensive).
- Cost Efficiency: Fable5.1 at low level is as effective as Fable5 at high level, but at a third of the cost. The price for caching and reading has been significantly reduced, making long-term use more cost-effective.
- Flexibility: You can increase the model's “effort” temporarily for complex tasks and then reduce it back to save resources.
2. Model Behavior Changes: Just Update Your Prompts
Fable5.1 has seven behavior changes compared to Fable5, and Anthropic has provided “correction templates”:
- Reduced Parallel Tool Usage: The model now uses only one tool at a time, which may increase processing time. Solution: Have the model list the required tools first and batch the non-repeated calls (e.g., checking the weather and stocks together).
- Fewer Progress Updates: The model might remain silent during tasks. You can use the `display:updates` prompt or add a prompt to inform the user of its progress and summarize the results.
- More Concise Writing Style: If the model’s output is too verbose, add a prompt to remove unnecessary formalities.
- Changed Formatting Preferences: If the model used to use lists or bold text, you may need to adjust your prompts accordingly (e.g., use “list items” instead of “bold text”).
- Avoid Interrupting the Process: If the model frequently asks if to continue, add a prompt indicating that the user can proceed independently.
3. Context Handling: Simplify History for Better Efficiency
Fable5.1 has stricter rules for handling conversation history:
- Historical Data is Read-Only: Changes to previous content can affect subsequent predictions. New instructions should be added as “temporary messages” that disappear after the next user input.
- Delayed Compression: Since caching is now cheaper, you don’t need to compress historical data immediately. You can compress it later, especially for longer conversations.
- Focus on Important Information: When compressing, retain the user’s original words, difficult solutions, and decision-making points.
4. Cleaning Up Old Prompts
Anthropic has removed 80% of the redundant prompts from Fable5, and Fable5.1 continues this effort:
- Remove Unnecessary Prompts: Verification requests, repeated instructions, outdated examples, and contradictory rules.
- Automatic Prompt Auditing: Use the `claude-code` or `claude-api prompt-audit` tool to check for redundant or contradictory prompts.
5. Small Issues and Solutions
- Low-Level Behavior: The low-level setting may not include search functionality (which could be outdated). Solution: Use a higher level or add a prompt reminding the model to search before answering.
- Security Issues: Certain questions may trigger false positives. Adjust prompts accordingly (e.g., change “Can it compile?” to “Are there any bugs?”).
- File Optimization: Avoid rewriting entire files for minor code changes. Add a prompt to modify only the relevant parts.
Conclusion
Fable5.1 is designed to be more intelligent and cost-effective. Developers can achieve better results with fewer resources by following the official prompts. For end-users, this means faster responses and more tailored services, thanks to the improved cost optimization.