NemoClaw Knowledge Wiki

Compare of Claude Opus 4.5 vs ChatGPT 5.2 Matt Maher

4 min read

https://www.youtube.com/watch?v=iUzrE3-FHgA Here is a summary of the video comparing OpenAI’s new GPT-5.2 models against Anthropic’s Claude Opus 4.5 using a massive, complex “One-Shot Build” benchmark.

The Experiment: The “Impossible” PRD

Instead of standard benchmarks, the host (Matt) created a massive Product Requirements Document (PRD) for a movie/TV companion app called “Showbiz.”

  • The Input: A huge folder of documentation including technical specs, design tokens, personality guidelines, and complex feature logic (like “Alchemy” concept blending).
  • The Goal: Ask the models to read the docs and build the entire application in one shot.
  • The Models Tested:
    1. GPT-5.1 Codex Max Extra High (Previous State of the Art).
    2. GPT-5.2 Medium (The new “recommended” efficiency model).
    3. GPT-5.2 Extra High (The new reasoning-heavy flagship).
    4. Claude Opus 4.5 (Anthropic’s powerhouse).

Phase 1: The Initial Build (One Shot)

The models were given 2 hours (simulated) to build.

  • GPT-5.1 (Previous SOTA): Failed significantly. The search functionality 404’d, the UI was basic, and it missed the vast majority of requested features.
  • GPT-5.2 Medium: A major step up. It built a functional UI with working search and basic details. It missed visual elements like “Key Art” but nailed the logic for features like “The Scoop.”
  • GPT-5.2 Extra High: Produced a very elegant, polished UI. It included interactive search and grouped recommendations upfront. However, it initially missed specific data points like show trailers and cast images.
  • Claude Opus 4.5: The most feature-complete on the first try. It included seasons, episode lists, inline playable trailers, and complex metadata. However, the UI was messy and visually cluttered compared to GPT-5.2.

Initial Verdict: Opus won on logic/completeness; GPT-5.2 xHigh won on UI/UX and polish.

Phase 2: The “Delta” Fix (The Game Changer)

Matt performed a quick “Gap Analysis”—he simply asked the models to compare what they built against the original PRD and fix what they missed. This resulted in near-production quality apps.

GPT-5.2 Extra High (Final Result)

  • Visuals: Extremely polished, professional UI.
  • Data Visualization: Created complex charts for actors (e.g., Ryan Reynolds’ career analytics, box office ROI, genre breakdowns).
  • Logic: The “Alchemy” feature (blending two movies to find a third based on concepts) worked perfectly.
  • Completeness: It recovered almost all missing features, creating a 95% complete application.

Claude Opus 4.5 (Final Result)

  • Visuals: Fixed the CSS breaking issues, resulting in a dense but usable interface.
  • Features: Retained its lead on specific media features like inline video players.
  • Personality: The “Ask” (Chat) feature felt more distinct and aligned with the “witty friend” persona requested in the PRD.

Key Takeaways & The “Communication Gap”

  1. The “Gap Analysis” Strategy: The models cannot do a massive build perfectly in one shot yet. However, if you ask them to critique their own work against the requirements and apply fixes, they can achieve 90-95% completion extremely fast.
  2. Productivity Leap: What previously took Matt nearly a year to build manually was recreated by these models in roughly 2 hours.
  3. The Communication Difference (Vital Insight):
    • Anthropic (Claude/Opus): Matt prefers working with Claude because it communicates. It repeats back what it heard, explains its plan, and confirms its understanding before coding. This creates a “pair programmer” feeling.
    • OpenAI (GPT): GPT is a “silent executor.” You give it a prompt, and it immediately starts coding without explaining its interpretation. While the code is excellent, it requires the user to have absolute trust in the model, which can be unsettling when errors occur.

Final Conclusion: Both GPT-5.2 and Opus 4.5 are massive leaps forward. While neither is perfect instantly, the ability to build sophisticated, complex applications with just a “Build” pass and a “Fix” pass has fundamentally changed software development.

Graph View