GPT-5.2 is a total monster

GPT-5.2: Detailed Review & Performance Analysis

Key Concepts:

• GPT-5.2: OpenAI’s latest and most capable AI model, available on paid ChatGPT plans.
• Gemini 3 Pro: A leading competitor to GPT-5.2, developed by Google.
• Multimodality: The ability of an AI model to process and understand multiple types of data (text, images, code).
• Context Window: The amount of text an AI model can process at once (GPT-5.2: 400k tokens, Gemini 3: 1M tokens).
• Hallucination Rate: The frequency with which an AI model generates factually incorrect or nonsensical information.
• Benchmarks: Standardized tests used to evaluate AI model performance (e.g., GBT Val, Swebench Pro, ARC AGI2, OCR).
• Agentic Coding: The ability of an AI to autonomously write, debug, and execute code.
• Ray Tracing: A rendering technique used to create realistic 3D graphics.
• Tool Use: The ability of an AI to utilize external tools (e.g., Python interpreters) to enhance its capabilities.

---

I. Initial Demos & Coding Prowess

The video begins with demonstrations of GPT-5.2’s capabilities, focusing on tasks beyond simple content generation (emails, blog posts). The emphasis is on challenging prompts involving complex coding, reasoning, and visual analysis.

• Beehive Simulation: GPT-5.2 successfully generated a functional HTML simulation of a beehive, complete with hexagonal cells, worker bee paths, honey storage, and adjustable parameters (colony size, resource availability). Notably, the simulation’s design – bees entering from a single entrance – was more biologically accurate than outputs from Gemini 3 Pro and previous GPT-5 versions. The code generation time was 19 seconds.
• Photoshop Clone: GPT-5.2 created a surprisingly comprehensive clone of Photoshop in HTML, including brushes, layers, edit history, filters, and blending options. The preview function allowed for real-time interaction with the generated application. All core features (brush size, color, opacity, layer manipulation, filters like grayscale, invert, blur, sharpen, edge detect, brightness, contrast, hue rotate, saturation) functioned correctly. This was significantly more functional than previous attempts with Gemini 3 and GPT-5. Code generation time was also 19 seconds.
• 3D Scene Generation: When prompted to create a 3D scene from an image using Three.js, GPT-5.2 produced a detailed render comparable to Gemini 3 Pro, with slightly more detail in the cherry blossoms.

II. Advanced Coding Challenges & Performance

The video then showcases GPT-5.2 tackling even more complex coding tasks.

• Two-Sphere Ray Tracing Simulation: GPT-5.2 successfully generated a simulation of two metallic spheres suspended above a street scene, a task previously impossible for other models. The code incorporated ray tracing with accumulation in WebGL 2 and utilized publicly available 3D street panoramas. Crucially, the spheres accurately reflected each other, demonstrating advanced rendering capabilities. The model’s thought process included searching for street panoramas and implementing ray tracing pseudo-code.
• Windows 11 Clone: GPT-5.2 created a functional clone of the Windows 11 desktop, including a start menu and working applications (MS Word, Excel, PowerPoint). While the visual fidelity wasn’t perfect, the core functionality of the applications was present. Excel and PowerPoint were fully functional, surpassing the capabilities of Gemini 3 Pro in this regard. Code generation time was 10 seconds.
• Interactive 3D Night Sky Viewer: GPT-5.2 generated an interactive 3D night sky viewer with labeled constellations in a single HTML file. The viewer included adjustable parameters for star size, label size, and line opacity.
• "Find Waldo" Challenge: GPT-5.2 attempted to locate Waldo in a complex image, a task that previously stumped other models. While it took 13 minutes to process, it ultimately identified and circled Waldo correctly, demonstrating advanced image analysis and pattern recognition.

III. Multimodal Capabilities & OCR Performance

The video demonstrates GPT-5.2’s ability to process and understand images.

• Demon Slayer Character Labeling: When presented with an image of multiple Demon Slayer characters, GPT-5.2 accurately identified and labeled each character with bounding boxes.
• Table to Spreadsheet Conversion: GPT-5.2 successfully converted a complex table (with nested columns and missing cells) into a functional Excel spreadsheet.
• Flowchart to Interactive Canvas: GPT-5.2 generated an interactive canvas representation of a flowchart, allowing for node manipulation and zooming. However, the arrow connections weren’t entirely accurate.
• Medical Image Analysis: GPT-5.2’s performance on identifying lesions in medical images was less impressive, with several incorrect identifications.

IV. Benchmarks, Specs & Comparison with Gemini 3 Pro

The video delves into the technical specifications and benchmark results of GPT-5.2.

• GPT Val Benchmark: GPT-5.2 (Pro & Thinking) is the first model to consistently outperform expert-level human workers on real-world tasks spanning multiple industries.
• Swebench Pro vs. Verified: OpenAI’s preference for Swebench Pro (testing four languages) over Swebench Verified (Python only) is highlighted as potentially biased, as GPT-5.2 performs better on Pro.
• Reasoning & Math Benchmarks: GPT-5.2 excels in GPQA Diamond, Frontier Math, and Competitive Math.
• ARC AGI2: GPT-5.2 achieves a score of 52.9% on ARC AGI2, demonstrating strong pattern learning capabilities.
• Context Window: GPT-5.2 has a context window of 400,000 tokens (300,000 words), while Gemini 3 Pro has 1 million tokens.
• Knowledge Cutoff: GPT-5.2’s knowledge cutoff is August 2025, more recent than many competitors.
• Pricing: GPT-5.2 is priced at $4.8 per million tokens, slightly more expensive than Gemini 3 Pro but cheaper than Anthropic’s Opus 4.5.
• Independent Leaderboards: GPT-5.2 (Extra High) is generally comparable to Gemini 3 Pro on independent leaderboards like Artificial Analysis. However, performance varies across different benchmarks (SimpleBench, OCR).
• Hallucination Rate: GPT-5.2 has a hallucination rate of 78%, lower than Gemini 3 Pro (88%) but higher than some other models.

V. Conclusion & Key Takeaways

GPT-5.2 represents a significant advancement in AI capabilities, particularly in coding and complex reasoning. While generally comparable to Gemini 3 Pro in overall performance, it excels in specific areas like agentic coding and certain reasoning benchmarks. The model’s multimodal capabilities and ability to utilize external tools further enhance its functionality. However, it’s important to consider its limitations, such as the smaller context window compared to Gemini 3 Pro and the potential for hallucinations. The video emphasizes the rapid pace of AI development and the challenges of objectively evaluating model performance.

Quote: "This is the most thorough Photoshop clone that I've seen so far. Much better than Gemini 3, which looks like this. Some of the features don't really work." - Demonstrating GPT-5.2's superior coding capabilities.