Claude Fable 5 prompt word leaked and the effect was measured in 6 hours. Is it crazy?

Fable 5 has just been launched, and the system prompts have been leaked:

After looking at these prompts, there are a few key points to note:

• First, Fable has added a persistent storage API (window.storage) for Artifacts. Artifacts are the unique content generated by Claude's code, such as HTML pages and React components. Previously, Artifact data could not be saved and was more like a one-time demo. Now, the data can be saved across sessions, enabling the creation of "memory-based" tools such as leaderboards, time trackers, and diaries.

• Second, Fable has improved the complete logic of the MCP App connector. If you want to connect to external services (such as ordering, taxis, music, etc.), Fable will first check the available service catalog and then present the options for your confirmation. Especially for services that incur costs, it has implemented a detailed set of opt-in requirements.

Additionally:

• Third, when Fable detects requests related to network security, biochemistry, or distillation, it will automatically send the response to the secondary model Opus 4.8 for processing and notify the user of the downgrade.

• Fourthly, Fable has added a new feature called `long_conversation_reminder` (a reminder for long conversations). When a conversation becomes too lengthy, a reminder is added to the end of the user's information to alert the model not to forget the original conversation, as it has lasted for an extended period of time.

Repository link:

https://github.com/elder-plinius/CL4R1T4S/blob/main/ANTHROPIC/CLAUDE-FABLE-5.md

Although the civilian downgrade of Mythos was anticipated, it is still evident from the system's prompts that Anthropic places great emphasis on security.

It is worth mentioning that the mental health section occupies a considerable amount of space within the behavioral guidelines for Fable. Considering the suicide-related lawsuits involving chatbots in the past one or two years, it is not difficult to understand why this part is described in such detail...

The original text appears to be a sequence of characters or tokens that do not form a coherent sentence or piece of text in Chinese. It may be a code snippet, a set of commands, or some other type of data. Since the text does not contain any meaningful Chinese words or phrases, it cannot be translated into English. If you provide additional context or information about the purpose of this text, I may be able to assist with the translation.

The translation is as follows:

Claude is concerned about people's physical and mental health, and avoids encouraging or promoting self-destructive behaviors such as addiction, self-harm, eating or exercise disorders/unhealthy practices, as well as highly negative self-talk or self-criticism.

Even when users explicitly request it, Claude can still avoid creating content that might support or exacerbate self-destruction.

When discussing "danger reduction strategies" or "safety plans" with people who have suicidal thoughts or self-harm impulses, Claude's approach is not to name, list, or describe these things specifically. Even when it's necessary to inform them about what they should avoid, Claude avoids mentioning them directly, as doing so might inadvertently trigger these individuals.

So, let's take a look at the actual test results for Fable 5.

This is a demo created by a netizen using Fable 5 and the Ancient Scrolls feature:

Fable uses the 3D forest scene built by ThreeJ:

Possessing a space simulation with 5000 objects:

New York Skyline Demonstration:

The Gothic city was swallowed by the waves:

Netizen Victor Taelin used a real project for the testing.

HVM5It is an underlying system related to high-performance computing. He has used it before; it has 32 cores.GPT-5The agent ran for approximately 20 hours. Although it could achieve a maximum acceleration of 2 times, the quality of the code deteriorated due to code bloat. Later, it was allowed to use Opus 4.8.GPT-5.5Optimization took 8 hours, with Opus providing an effective speed increase of 6% to 34%. GPT achieved even better results, but the files produced by GPT cannot be used.

As a result, Fable 5 only took 2 hours to complete the process, and it achieved a 1770% improvement on one benchmark, while the other 4 benchmarks showed an increase of more than 100% on average, with an additional 22%.

The first reaction of the netizens was disbelief; they suspected that the benchmark might be "hardcoded," as they had previously been deceived by similar issues with GPT.

After reading the explanation, he realized that the optimization direction for Fable was indeed reasonable.

Because when HVM5 processes dynamic pattern-match nodes, many unnecessary branches are also sent to garbage collection, which wastes a lot of time. Previously, only static matches were optimized, while dynamic matches were not. Fable has identified this issue, which is why the test results are becoming quite exaggerated.

This is the root cause analysis of the HVM5 bug provided by Fable 5:

Turkish internet user Alican Kiraz conducted a series of comparison tests between Fable 5 and GPT 5.5.

His conclusion is that Fable 5 cost a lot of money to run, 360.55 US dollars; GPT-5.5 only cost 6 US dollars. However, when it comes to detailed optimizations, Fable 5 has indeed made optimizations that are more fundamental, more hardcore, and closer to the mindset of performance engineers.

These are his test conclusions:

According to real task tests using the third-party programming tool Augment Code, Fable 5 has truly strong programming capabilities.

The tests were run together for 489 programming tasks, and Fable 5 clearly outperformed in terms of overall performance and accuracy, with a total score of +0.224 and an accuracy score of +0.191.

However, there are also examples of disappointing results from actual tests. For instance, internet user Ryan R. Hughes submitted a large PR (Pull Request) containing 74 files for review to Fable 5. The review process took over 34 minutes and consumed 42% of the available Claude Code session time, yet only 16 issues were identified.

Such examples are not uncommon; Fable 5 is just a bit too expensive.

Some people believe that, based on actual tests, Fable 5 has a weak ability to parse long texts:

More negative reviews focus on the model degradation process:

The current API price for Fable 5 is $10 per million input tokens and $50 per million output tokens.

In a side-by-side comparison, this price is approximately twice that of Opus 4.8 and more than three times that of Sonnet 4.6. The actual cost of a single complex workflow is significantly higher than that of any previous generation of Claude models.

And its sensitivity has also been increased: the overall design of the guardrail has become more conservative; they prefer to potentially injure someone accidentally than to prioritize safety first.

Regarding the issue of high costs, I have collected the following practical tests from netizens that can help everyone save money while still achieving good results: a model mixing solution:

In IDEs like Cursor that support hybrid models, different models can be used on the platform in a phased manner:

Phase 1: Code review and preliminary project analysis, used to warm up the context and understand the project structure.GPT-5.5High, MiniMaxM3、KimiK2.6Either Composer 2.5 or one of its later versions.

Phase 2: Develop a project plan using GPT-5.5 Very High or Opus 4.8 to prepare the project plan.

Stage 3: Analyze the use of the project plan. Review the Fable 5 project plan to identify potential errors. If modifications to the plan are required, use GPT-5.5 with a Very High priority. Stage 3: Analyze the use of the project plan. Review the Fable 5 project plan to identify potential errors. If modifications to the plan are needed, use GPT-5.5 with a Very High priority.

Stage 4: Implementation and use of the project plan, utilizing tools such as MiniMax M3, DeepSeek V4, Max, Composer 2.5, or Sonnet 4.6 to carry out the implementation plan.

Stage 5: Final Review Use the GPT-5.5 code review results to check whether the implementation meets the original plan.

Using the most expensive resources in the most appropriate places is the right way to approach Fable 5.

The effectiveness of killing the mad in Claude Fable 5 is a matter of personal opinion; some may agree, while others may disagree.

But with it, the token consumption rate is very clear – it’s absolutely crazy.

The article is sourced from the WeChat public account “JackCui”, written by “JackCui”.