I Made GPT-6 Astra & Fable 5.1 Build the Same App (RAW RESULTS)

I Made GPT-6 Astra & Fable 5.1 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 👥 24K 📅 September 5, 2026 ⏱ 26 min 👁 274K 📄 science communication 🧭 2026-09-10
Available in: English (current) Français

Keywords

AI modelscode generationLLM performancecost analysis3D web apps

Summary

In this 26-minute video, the creator Pat Simmons pits two AI coding agents, GPT-6 Astra and Claude Fable 5.1, against each other with identical one-shot prompts to build three different web applications: a 3D iPod loaded with early 2000s songs, a scrubbable 3D D-Day reconstruction, and a Call of Duty-style Nuketown clone. Both models received the same prompt files and were given no revisions before live deployment. The video documents the build times, costs, and quality of each output. For the iPod, Astra finished in 37 minutes and cost $13, while Fable took 1h20 and cost $41; Astra’s build was visually superior, though Fable included extra features like sound effects. The D-Day build had a stark contrast: Fable completed in 40 minutes, but Astra took 10h51, spending hours on historical research and terrain accuracy. Astra’s recreation featured detailed 3D troops and towns, albeit with modern-looking soldiers, while Fable’s was more map-like. For the Nuketown clone, both took ~47 minutes, with Astra producing more realistic graphics and player models. Overall, Astra consistently delivered higher-quality results at lower cost, except for the D-Day build’s time. The video provides transparent cost breakdowns and practical insights into AI-generated coding, but it remains a single informal test.

202 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video’s value lies in its practical, head-to-head comparison with real-world tasks and transparent cost/time metrics. The creator clearly demonstrates each model’s strengths and weaknesses, providing tangible evidence for claims. The argumentation is well-supported by live demonstrations and usage statistics, making it both informative and persuasive. However, the test lacks scientific rigor: it is a single instance, with no repetition, and the creator’s subjective evaluation of graphics and playability introduces bias. Despite this, the video serves as a valuable field test for users deciding between these AI tools.

Scientific Rigor, Source Quality, Title Accuracy

The video references only the creator’s own blog and newsletter, with no external sources cited. The creator mentions inspiration from other builders (Pro Shirano, Dan Shipper) but without links. The title is slightly misleading as it says ’the same app’ while showing three builds; nevertheless, it accurately captures the ‘raw results’ nature. The scientific rigor is moderate: the methodology is transparent, but the absence of controlled conditions and peer validation limits its reliability. The title–content gap is minor and does not significantly affect the overall assessment.

189 words

Title / Content Match

The title suggests building one app, but three different apps are built; however, the 'RAW RESULTS' framing is appropriate.

Quality & Reliability

7/10

The video provides a transparent comparative test of two AI coding models with actual costs and build times, but it is a single anecdotal test without controlled conditions or external verification.

Chapters

Cited Sources

Contribution & Novelties

This video contributes a practical, side-by-side evaluation of two frontier AI coding agents, highlighting differences in cost, speed, and output quality. It goes beyond simple demos by providing real usage costs and time-to-completion, which are critical for developer adoption. The D-Day build showcases an unexpected depth of research and detail in Astra’s process, including tide calculations and terrain verification, offering a glimpse into agent reasoning. The Nuketown build demonstrates the gap in visual fidelity between the models.

Pour aller plus loin :

110 words

Radar Profile

The radar profile shows high scores in information quantity (9) and technical depth (8), but lower reliability (7) and information quality (7). This indicates a comprehensive and detailed practical comparison that lacks formal experimental rigor, typical of a motivated field test.

Reliability 7/10

💬 Based on the 30 comments analyzed, the overall sentiment is positive, with viewers praising the comparative format and the detailed results. Some commenters note that the video feels like an advertisement for OpenAI, but the majority are enthusiastic about the practical insights and cost breakdowns.