I Made GPT 5.6 and Fable 5 Build the Same App (RAW RESULTS)

I Made GPT 5.6 and Fable 5 Build the Same App (RAW RESULTS)

by Pat Simmons · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Pat Simmons 👥 25K 📅 July 11, 2026 ⏱ 26 min 👁 119K 📄 original study 🧭 2026-09-14
Available in: English (current) Français

Keywords

GPT-5.6Fable 5app cloningMSCHFlearning platform

Summary

In this video, Pat Simmons performs a rigorous head-to-head comparison of OpenAI’s GPT-5.6 (alias Sol) and Anthropic’s Fable 5 on three identical one-shot app-building tasks. The first task is a pixel-for-pixel clone of the mockup generator shots.so, where GPT-5.6 takes nearly three hours but delivers an exceptionally faithful and functional replica, while Fable finishes in about 30 minutes with a noticeably inferior result. The second task challenges the models to invent a viral MSCHF-style drop and build its website. GPT-5.6 produces a conceptually complex and interactive ‘Human Resources’ satire site, whereas Fable’s ‘Free Shirt’ concept is simpler and less polished. The third task asks for an interactive NYC learning platform; both models initially deliver static websites, but after receiving feedback, they generate impressive 3D interactive maps. However, GPT-5.6 again edges out Fable in design and interactivity. Throughout the video, Pat provides detailed cost breakdowns, shares the exact prompts used, and offers his skill repository for cloning apps. The overall conclusion favors GPT-5.6 for design fidelity, creativity, and depth, though Fable is faster and also capable. The video is well-structured, transparent, and offers practical insights for users of these AI models.

190 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video’s value lies in its direct, controlled comparison of two leading AI models under identical conditions, using realistic, complex tasks that go beyond simple coding benchmarks. The argumentation is solid: Pat clearly defines each task, states the rules (one-shot, no revisions, deployed live), and provides visual evidence of the results. He acknowledges limitations, such as the single-run nature and potential performance variability, and even gives both models a second chance on the third task to test feedback handling. The reasoning is inductive and data-driven, though subjective elements like ‘design quality’ and ‘creativity’ are evaluated qualitatively. The cost breakdown adds practical value for users mindful of token expenditure. Overall, the information is highly useful for AI practitioners and enthusiasts, and the methodology, while not exhaustive, is robust enough to support the conclusions drawn.

Scientific Rigor, Source Quality, Title Accuracy

The rigor is commendable: identical prompts are used, processes are documented, and full code is shared via the GitHub repository. The sources cited are the provided description links: the clone-app skill repo, the bootcamp, and the newsletter, which are relevant and offer additional resources. However, no external scientific literature is referenced; the video relies on empirical testing rather than established benchmarks. The title matches the content precisely, as it presents unfiltered results without overhyping. The evaluation criteria are transparent, though some might argue that a single iteration is insufficient for definitive conclusions, a point also raised in the comments. Overall, the video maintains a high standard of transparency and methodological clarity.

259 words

Title / Content Match

The title accurately reflects the content: a direct, unedited comparison of two frontier AI models building the same app, with results presented raw and unfiltered.

Quality & Reliability

8/10

The comparison is methodical with identical prompts and transparent sharing of all code and prompts, but relies on single-run outputs, limiting generalizability and statistical robustness.

Chapters

Cited Sources

Contribution & Novelties

The video contributes a practical, hands-on comparison of two frontier AI models on realistic app-building tasks, offering transparency via shared prompts, code, and cost data. It highlights GPT-5.6’s superiority in design fidelity and interactivity, while noting Fable’s speed. The methodology, though based on single runs, provides a baseline for further tests.

Pour aller plus loin :

  • AI alignment — Relevant to the satirical ‘Human Resources’ concept about proving humanity and AI’s role.
  • Prompt engineering — The video’s core technique in instructing models effectively.
  • Benchmarking — Context for how AI models are typically evaluated, compared to the ad-hoc tests here.

99 words

Radar Profile

The profile shows high scores in information quantity and technical level, reflecting the dense, practical content. Quality and reliability are slightly lower, due to subjective evaluation and lack of replication. Overall, a well-rounded, informative resource with minor methodological caveats.

Reliability 7/10

💬 positif — Sur les 30 commentaires analysés, la grande majorité exprime gratitude et enthousiasme pour la comparaison approfondie, avec quelques remarques constructives sur la nécessité de tests multi-pass pour plus de robustesse.