ChatGPT Images 2.0 tops Gemini tests

AI image models turn text prompts into pictures, and the hardest part has been getting them to draw readable words, stable layouts, and repeatable characters. OpenAI and Google are now competing on exactly those tasks. (openai.com) (ai.google.dev) ZDNET tested OpenAI’s ChatGPT Images 2.0 against Google’s Gemini Nano Banana across nine prompts and gave the edge to OpenAI’s model. The article said ChatGPT Images 2.0 produced stronger realism and handled a wider range of styles more reliably. (zdnet.com) OpenAI introduced ChatGPT Images 2.0 and the API model gpt-image-2 on April 21, 2026. The company said the model improves text rendering, multilingual output, editing, layouts, and instruction following, with resolutions up to 2K. (openai.com) (community.openai.com) OpenAI also said gpt-image-2 added a “Thinking mode” for reasoning-heavy workflows, where the system can research, transform inputs, generate variations, and self-check before returning an image. The launch post said developers can use it in the API and in Codex. (community.openai.com) Google’s competing line is now split between Nano Banana 2 in the Gemini app and Gemini 2.5 Flash Image in the API. Google describes Nano Banana as its high-speed image model for conversational editing, low-latency generation, and multimodal workflows. (workspaceupdates.googleblog.com) (ai.google.dev) Google rolled out Nano Banana 2 on February 26, 2026, saying it replaced the earlier Nano Banana model for Gemini users. The company said paid users can generate 2K images, while the model is built to preserve up to five characters and ten objects in a single workflow. (workspaceupdates.googleblog.com) Third-party rankings also show OpenAI moving ahead, though the exact gap depends on the benchmark. OpenAI’s developer post said gpt-image-2 hit No. 1 across Image Arena leaderboards with a +242 lead in text-to-image, while Artificial Analysis currently lists GPT Image 2 at 1,332 Elo and Google’s Nano Banana 2 at 1,262. (community.openai.com) (artificialanalysis.ai) That difference in numbers reflects different scoreboards, not necessarily a contradiction. OpenAI cited Arena results from launch week, while Artificial Analysis publishes its own live leaderboard with separate sampling and scoring. (community.openai.com) (artificialanalysis.ai) The practical fight is less about pretty pictures than replacing design cleanup. OpenAI is pitching posters, charts, comics, slides, and multilingual graphics, while Google is pitching fast editing, real-world knowledge from Gemini, and production-friendly image generation inside Workspace and the Gemini app. (openai.com) (workspaceupdates.googleblog.com) ZDNET’s test adds one more public data point to that race: OpenAI’s newest model is winning more head-to-head comparisons right after launch, and Google is still competing on speed, app integration, and editing workflows. (zdnet.com) (ai.google.dev)