Summary of Key Points
Anthropic's AI model, Fable 5, has been restored after being unavailable for over half a month, but its performance is far from what it used to be: security filters have become excessively strict (even asking questions like "Does a bumblebee fart?" may result in rejection), and it frequently reverts to the older Opus 4.8 model. The token consumption rate is incredibly high, leading to a significant decline in user experience. Nevertheless, Fable 5 still possesses strong capabilities, making it suitable for complex engineering tasks and code reviews. Both officials and users have provided some tips to minimize costs and maximize efficiency when using it. Additionally, the Claude desktop version has released a Linux test edition, expanding its range of applications.
1. The Reason for the Restoration: A Vulnerability Report and Enhanced Security Measures
The direct cause of Fable 5's suspension was a discovery by Amazon researchers that specific prompts could bypass its security mechanisms, allowing the model to generate code that could exploit software vulnerabilities. Although Anthropic argued that the attack was not on par with that of Mythos 5 and cited competitors like GPT-5.5 and Kimi K2.7 as being able to recognize such vulnerabilities as well, they decided to urgently shut down Fable 5.
After restoration, Anthropic retrained its security classifier, which now blocks 99% of potential vulnerability exploitation attempts. However, this has come with significant drawbacks:
- The model prioritizes blocking attacks related to cybersecurity (vulnerability exploits, malware), biological/life science content (experimental methods), and requests aimed at extracting the model's internal thoughts;
- Even legitimate tasks are sometimes misclassified: programmers asking about standard coding issues, doctors and scientists exchanging greetings, or even questions like "Does a bumblebee fart?" may all result in Fable 5 being redirected to Opus 4.8.
Anthropic acknowledges that the biological/chemical classification algorithm has not been updated and covers too broad a range of topics; improvements are still in progress, so users in these fields will have to wait.
2. User Complaints: Spending More for a Lower-Quality Model
Fable 5 costs twice as much as Opus 4.8 (10 USD per million tokens input, compared to 50 USD for Opus 4.8), but users often receive the same results from the older model:
- Some users spent $90 on three complex prompts and used up 60% of their token quota in just 30 minutes;
- A simple question could consume 2% of a Pro plan's token quota in 5 hours;
- Even worse, users who pay for Fable 5 may end up using Opus 4.8 due to security misclassifications, effectively getting a lower-quality service for a higher price.
The community is full of complaints about the high consumption rate, lack of improvement in output quality, and frequent rejections to older models.
3. The Value of Fable 5's Strengths
Despite the poor user experience, Fable 5 still has significant strengths:
- Complex Engineering: It can complete system implementations that require multiple days of iteration, such as creating a clone of "Rocket League" or implementing Spider-Man swinging games in Godot (with fewer errors than GPT-5.5);
- Visual Understanding: It can accurately analyze complex technical diagrams and web page screenshots and use tools to process blurry images;
- Enterprise Tasks: It performs better in financial analysis, table and slide processing;
- Code Skills: It has a higher bug detection rate compared to Opus 4.8 and can search code repositories and historical commits;
- Multi-Agent Collaboration: It is adept at using multiple agents to handle subtasks, making it suitable for long-term autonomous projects.
Anthropic also states that Fable 5 consistently ranks first on the CursorBench benchmark tests (although it comes at a higher cost).
4. How to Get the Most Out of Fable 5: Official and User Tips
To make Fable 5 more cost-effective, both officials and users have provided the following tips:
- Official Recommendations:
1. Prevent the model from showing its internal thinking (which can trigger security filters); use tools to display progress at key stages;
2. Define clear task boundaries (e.g., avoid requesting additional features or modifying code);
3. Verify results for long-term tasks (acknowledge failures and don't skip steps);
4. Create an external reference system using Markdown to save costs on repeated queries;
5. Use sub-agents efficiently: assign smaller tasks to auxiliary agents while the main model handles the overall project.
- User-Saved Tips:
1. Use Fable 5 only for complex tasks and avoid higher-tier settings (which are more token-intensive with limited performance improvements);
2. Communicating in Chinese saves tokens (tests show that Chinese usage results in almost zero additional token consumption, compared to 1.4 times more for English);
3. Outsource high-resource-consuming tasks (e.g., code repository analysis) to other models and then feed the results to Fable 5.
5. An Update on Claude: Finally Reaching Linux Users
In addition to Fable 5, the Claude desktop version has been released for Linux (Ubuntu 22.04 and Debian 12), with features such as Chat, Cowork, Code tabs, parallel sessions, visual diff reviews, and an integrated terminal. However, it currently lacks screen control and voice input, and global shortcut keys depend on the desktop environment. Fedora and RHEL are not yet supported.
This move brings Claude closer to Linux developers' workspaces, but it also means more security restrictions. Users must balance the model's capabilities with the additional constraints.
In Conclusion
Fable 5 is a powerful model, but its strict security measures and high cost make it less practical for everyday use. If you decide to use it, you'll need to learn how to optimize its performance by adjusting prompts according to the guidelines provided and use it wisely in suitable contexts. While Anthropic aims to balance security and user experience, the current setup seems to prioritize security at the expense of convenience.