GPT 5.2 and Image-gen-2 from Open AI - A final swing at Google?

OpenAI’s latest push—GPT 5.2 plus an image model billed as “Image-gen-2”—is landing as a serious, if uneven, challenge to Google’s top generators. Early community tests suggest the image system can produce more scientifically accurate labeled diagrams than Google’s “Nano Banana Pro,” while GPT 5.2’s coding and 3D demos look competitive with Google’s best models—especially in how much complete functionality can be generated from code-only prompts.

On the image side, a human-cell diagram prompt (“Create a fully labeled diagram of a human cell with at least 10 elements”) becomes the main battleground. The diagram produced by “Hazel Gen 2” (widely assumed to be OpenAI’s GPT Image 2) gets several core labels right—human cell spelling, the plasma membrane placement and detail, and multiple organelles such as the cytoskeleton, centrosome, lysosome, mitochondria, Golgi apparatus, vesicles, ribosomes, and smooth/rough endoplasmic reticulum. But accuracy breaks down in specific scientific spots: the Golgi apparatus appears scrambled or mislocated, ribosomes are rendered as tiny dots, and some organelle associations look swapped or misplaced (including confusion between vesicles and mitochondria, and a lysosome that points toward the nucleus rather than its proper location). Even so, the tester’s bottom line is that Nano Banana Pro still wins in “scientific accuracy,” but by a smaller margin than expected.

A second prompt—a chai recipe with step-by-step instructions—shows fewer obvious errors and highlights a different kind of strength: GPT Image 2’s output includes more structured, step-aligned visuals and a more “textbook” feel, while Nano Banana Pro’s visuals lean more artistic and ingredient-focused. The comparison ultimately comes down to preference: one system is more semantically tidy and step-by-step, the other more visually stylized.

Then the focus shifts to GPT 5.2, released as the model’s rollout lands. In Element Arena demos, GPT 5.2 generates substantial 3D scenes and interactive projects from code prompts alone—no image inputs—ranging from a Golden Gate Bridge with adjustable weather, time of day, and traffic density to voxel foliage, animated fish, and even a rocket launch with particle effects. Benchmarks cited from LM Arena and AIM-style leaderboards place GPT 5.2 near the top: it’s described as improving over GPT 5.1, with “high thinking” variants trading off cost and accuracy against Claude 4.5 Opus and Gemini 3 Pro.

But practical tests also reveal rough edges. In a water-physics HTML challenge (interactive 3D with reflections and wave simulation), GPT 5.2 initially produces broken or incomplete code, then iterates toward a working version—though it introduces issues like “ghost lemons” and glitchy physics. In a physics-based jelly platformer, GPT 5.2 generates very large codebases (over a thousand lines) and playable-looking graphics, yet physics can become unbalanced or unplayable due to assumptions about frame rate and tuning. The overall verdict: GPT 5.2 is surprisingly competitive and strong as a starting point for real projects, while Google’s Gemini 3 Pro remains formidable—sometimes even better at producing tighter, more immediately functional results in specific physics tasks.