Claude Fable 5 Built This in Claude Code and I'm Blown Away
Leon van Zyl ·2026-06-10 ·2 min read
Summary written by us from the video's transcript. The video, and everything in it, is Leon van Zyl's work.
Learn how to use Anthropic's Fable 5 model in Claude Code for a complex agentic coding task—building a browser‑based game with reflections and ray tracing—and see the results compared to GPT 5.5.
Takeaways
- Fable 5 can handle highly complex agentic coding tasks like building a browser‑based game with advanced graphics, though it runs slower than earlier models.
- Choosing **extra high** reasoning effort balances quality and token usage; max effort may cause hallucinations.
- Claude Code’s built‑in **goal** command combined with feature‑file decomposition enables largely hands‑off implementation of large projects.
- Despite higher token costs ($10/$50 per million), Fable 5 delivers noticeably better visual results than GPT 5.5 for demanding graphics tasks.
- Using a cheaper model (Opus) for background review agents can reduce overall credit consumption without sacrificing final quality.
Introducing Fable 5 and Its Benchmarks
Anthropic’s new **Mythos** class model, Fable 5, outperforms Opus 4.8 on the SPEE benchmarks, scoring 80% versus 70% on Agented Coding Tasks.
The author sets a high‑bar test: instead of simple dashboards, he asks Fable 5 to create a full game engine with rendering, shaders, camera, and material systems that runs in a browser.
Setting Up Claude Code for Fable 5
Update Claude Code to the latest version, select **Fable** from the model selector (Opus 4.8 is default), and choose an appropriate reasoning effort. The author recommends **extra high**—high enough for complex tasks but avoiding the diminishing returns of the max setting.
Planning Phase and Cost Considerations
In planning mode, the prompt asks Fable 5 to produce a detailed implementation plan; the model takes about 15 minutes to generate it, indicating slower response times compared to earlier models.
Pricing: $10 per million input tokens and $50 per million output tokens—roughly double Opus 4.8’s cost—but potentially justified for harder reasoning tasks.
From Plan to Feature Files Using the Ramp Framework
The author exits planning mode, switches to edit/YOLO mode, and feeds the plan back to Fable 5 to expand it into a **detailed implementation plan** broken into individual feature files.
Claude’s built‑in **goal** command will later execute each feature automatically, with background agents reviewing the output (using Opus for review to save credits).
Running the Autopilot Implementation
A new repository is created; the goal command is issued: implement all features, run a background reviewer after each, and finish with a final review.
The process runs largely unattended, consuming a significant portion of the token budget (about 42 % of the allocated 20‑million‑token plan).
Result: Browser Game with Reflections & Ray Tracing
When execution finishes, the generated game demonstrates working reflections and ray tracing in the browser—mirrored panels rotate, lasers bounce, and an orb reflects its surroundings.
The author describes the visual quality as “stupidly impressive” and notes that the solution runs smoothly despite the complexity of the task.
Comparison with GPT 5.5
Running the same prompt on GPT 5.5 (using the goal command) took ~1 hour 20 minutes. The output shows functional ray tracing and shadows, but reflections are less accurate and overall visual fidelity is lower than Fable 5’s result.