Developers replicate Claude Fable 5 via leaked prompt

// 2h agoNEWS

Developers replicate Claude Fable 5 via leaked prompt

Following the U.S. government-mandated suspension of Claude Fable 5, the model's 1,585-line system prompt leaked online. Developers discovered they could recreate roughly 90% of the model's performance by injecting this leaked prompt into Claude Opus 4.8, creating a workaround dubbed 'Claude Fable 5 Lite'.

// ANALYSIS

System prompts are the new open-weights: you can't easily recall a model's alignment and steering behavior once its prompt is leaked, even if you pull the model offline.

* The leakage of the 1,585-line Fable 5 prompt demonstrates that model steering and behavioral personality are highly portable.

* Running Fable 5's prompt on Opus 4.8 showcases the resilience and capacity of older-generation frontier models to adapt to new agentic instruction sets.

* This workaround exposes the limitations of centralized API suspensions as safety or export controls, as prompt injection on existing models can bridge much of the performance gap.

// TAGS

anthropicclaudeclaude-fable-5claude-opus-4.8system-promptsafetyprompt-engineering

DISCOVERED

2h ago

2026-06-15

PUBLISHED

2h ago

2026-06-15

RELEVANCE

8/ 10

AUTHOR

codewithimanshu

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE6m ago

SmallHarness launches version 1.0 with model routing, followed by a hotfix introducing active effort routing to control reasoning depth for local and cloud LLMs.

SmallHarness, a terminal-based developer tool for running agentic LLMs on local hardware and cloud APIs, has officially released version 1.0.0, followed by a quick 1.0.1 update introducing active effort routing. The tool supports multiple backends—including Ollama, LM Studio, MLX, llama.cpp, and OpenRouter—and features safe execution of filesystem/shell commands via interactive approval gates and diff previews. The new active effort routing allows the system to analyze task complexity and automatically set reasoning effort levels (from minimal to max) sent to providers like OpenRouter and OpenAI, offering a highly customizable framework that optimizes performance, cost, and latency.

MODEL7m ago

User pans Kimi K2.7-Code in early testing

Early testing of Moonshot AI's Kimi K2.7-Code reveals rapid usage limit exhaustion and broken game generation. The user reports the model struggles with complex tasks, positioning competitor GLM 5.2 as a stronger alternative.

NEWS20m ago

A leak of the Claude Fable 5 system prompt allows users to replicate its frontier capabilities on Claude Opus 4.8 with high accuracy.

A social media post by CodeWithImanshu reveals that the system prompt for Anthropic's frontier Claude Fable 5 model has been leaked. When this extensive system prompt is run on Claude Opus 4.8, it reportedly achieves roughly 90% of the performance of the native Fable 5 model. This highlights how heavily Anthropic relies on prompt-level instruction tuning to define its new Mythos-class agentic behaviors.

Developers replicate Claude Fable 5 via leaked prompt