Summary
OpenAI’s GPT Image 2 launched on April 22, 2026, and Matthew Berman immediately put it through a gauntlet of real-world tests. The model debuted at the top of the LM Arena text-to-image leaderboard with an ELO score of 1,512 — a 250-plus point jump over the previous leader, Google’s Imagen 3 Flash (commonly called Nano Banana), which sat at 1,270. OpenAI describes GPT Image 2 as integrating ‘world knowledge’ and thinking-level intelligence directly into image generation, meaning the model draws on factual understanding to fill in contextual details without heavy prompting.
Berman tests a wide range of challenging prompts: a complete video game character sprite sheet covering dozens of animation states and genres, a hyperrealistic chalkboard with a multi-step algebra equation (which the model solves correctly when thinking mode is enabled), complex product shots, UI screenshot generation, and multi-step image editing sequences. Key findings include dramatically improved text rendering, stronger multi-image character consistency, and support for flexible aspect ratios from 3:1 to 1:3. Hand anatomy and granular style editing remain imperfect.
For developers and creators evaluating image generation tools in 2026, this video offers a practical prompt-by-prompt breakdown of where GPT Image 2 represents a genuine step change and where gaps still exist — useful for anyone considering it for game assets, product imagery, UI mockups, or content production workflows.
📺 Source: Matthew Berman · Published April 22, 2026
🏷️ Format: Review
