Claude Fable 5.2 is a freak
Use Fable 5.2 for difficult, tool-heavy creative or coding tasks; expect slow runs, high usage, and uneven results.
- For complex tasks, use Claude Code with local files and tools; the creator found it capable but slow and usage-intensive.
- For visual projects, add screenshot-based critic loops so an agent can identify issues and revise its work.
- Reserve Fable for tasks other models cannot handle: the review says it costs more, runs slower, and has uneven vision and research results.
The hands-on demos show how it handles complex tool-based creative tasks and where it fails. · Best for: AI practitioners evaluating agentic coding, creative workflows, and model trade-offs.
The creator tests Claude Fable 5.2 through agentic coding and software-control tasks, then compares its reported performance, cost, and limitations. Results are the creator's observations and benchmark claims presented in the video.
Model setup and a JavaScript explainer video0:00
- The creator recommends using Claude Code rather than chat for tasks involving local files, tools, and persistent projects.
- A JavaScript animation with Gemini TTS took about an hour and a half, including revisions; the creator liked the result.
- The video included background music even though the prompt asked only for narration.
Operating a virtual piano and producing a song4:03
- The model composed and played a piano solo through a browser interface; the creator said this test was already achievable with an earlier model.
- For a DAW project, it created instrumental tracks and multiple vocal parts using Voice Sauna, then arranged and mixed them.
- The music task hit the five-hour usage limit; the creator said the instrumental sounded strong but the vocals sounded synthetic.
Rebuilding a property and making a game12:31
- Using property photos and Blender MCP, the model made a detailed 3D reconstruction and camera tour; the creator found the layout mostly accurate.
- For an Unreal Engine game, it used a Sketchfab character and Mixamo animations, then iterated with a separate screenshot-review agent.
- The game took about seven hours and required usage-limit resets; the creator reported a GPU crash that the model then fixed.
Physics simulation, image recognition, and research19:58
- A from-scratch water-balloon simulation took about 35 minutes and included adjustable physics and lighting settings.
- It failed to find a frog in an image and misidentified other regions; the creator said DeepSeek found it.
- On six brain scans, the creator judged two answers correct; they found its chronic myeloid leukemia research response too concise.
Creative ideas, reported capabilities, and benchmarks24:26
- The creator found the animal-welfare ideas interesting but noted there was no single correct answer.
- The video reports gains in coding, reasoning, vision, robotics, and research, and describes company-reported math and physics work.
- The creator says the model is slow, expensive, and prone to making up answers when uncertain; it is also described as restricted for some research areas.
Cost, usage limits, and recommendation28:07
- The creator presents a comparison of over $26 per task for Fable, versus $10 for Opus and $6 for GPT-6 Astra.
- The creator says Fable is capped at 50% of weekly usage on paid plans and uses quota faster than Opus.
- Their recommendation is to reserve it for difficult reasoning tasks and avoid it for biology, cybersecurity, and AI research.
Step by step
- 1. Choose a difficult task that benefits from multiple tools, local files, or repeated revisions.
- 2. Run it through Claude Code and connect the required software or interfaces, such as Blender through Blender MCP.
- 3. Give the model clear output requirements and relevant references; provide tool documentation when needed, as the creator did for Gemini TTS.
- 4. For visual work, have a separate critic review screenshots, list issues, and send those issues back to the builder for another round.
- 5. Inspect the output yourself, correct errors, and account for long runtimes and usage-limit resets.
Tools & resources mentioned
- Claude Code — agentic coding with local projects and software tools.
- Gemini TTS documentation — instructions supplied to generate the explainer narration.
- Blender MCP — connection between the agent and Blender for 3D scene creation.
- Sketchfab — source of the character model used in the game prompt.
- Mixamo — source of character animation options.
- Voice Sauna — used to generate vocal tracks for the DAW project.
- RVC — mentioned as a possible way to improve synthetic-sounding vocals.
Notes are AI-generated from the video's transcript and may miss nuance.