ChatGPT Images 2.0 adds multilingual text and floor-plan output in creator tests

• OpenAI says ChatGPT Images 2.0 is built for composition-heavy jobs like comics, infographics, and floor plans, which matches the core evidence in the main HN thread and OpenAI's launch summary in the evidence pool.
• The biggest practical upgrade is text and layout control: OpenAI's launch summary says the model renders non-Latin scripts and keeps characters or objects consistent across multiple images, while the system card adds that thinking mode can use web search and generate multiple images from one prompt.
• Developers got the model on day one as gpt-image-2, and the HN discussion roundup zeroed in on pricing that stayed cheap at the low end even as high-quality renders got pricier.
• Early testers used it on prompt-heavy edge cases, including Simon Willison's raccoon-in-a-crowd prompt that the HN discussion roundup surfaced as a quick stress test for dense compositions and object placement.

OpenAI's own materials are worth skimming here: the launch post frames comics, infographics, and floor plans as the headline use cases, the developer announcement pitches gpt-image-2 as a production model for accurate, localized, on-brand assets, and the prompting guide quietly adds the practical bit: low quality is the latency play, while medium and high are for fidelity. Meanwhile Simon Willison's raccoon test and the main HN thread show where people immediately pushed it, crowded scenes, readable details, and prompts that used to fall apart.

OpenAI's launch framing is unusually specific for an image model. Instead of generic "better visuals" language, the company pointed at multi-paneled comics, infographic layouts, and floor plans in both the launch post and the developer announcement.

That lines up with what made the HN thread interesting. As the main thread summary puts it, commenters treated the release as a jump in composition-heavy output, not just prettier single images.

The most creator-relevant claim is not raw realism. It is readable text inside the image.

According to OpenAI's launch summary in the evidence pool, ChatGPT Images 2.0 improves text rendering across scripts including Hindi, Japanese, and Korean, and it can keep characters or objects consistent across multiple generated images. The system card adds the mechanism behind that pitch: thinking mode can reason through the prompt, call web search, and generate multiple images from a single prompt.

The developer announcement is even blunter about the intended workload. It calls out assets that need to be accurate, readable, localized, formatted for the destination surface, and usable without heavy cleanup.

The pricing chatter was more nuanced than "same price" or "price hike." In the HN discussion roundup, one commenter said API pricing looked mostly unchanged from the previous model, while another broke out the per-image estimates at 1024 by 1024: low quality moved from about $0.011 to $0.006, and high quality moved from about $0.167 to $0.211.

The official docs frame cost control around settings rather than a flat image fee. The prompting guide says gpt-image-2 is the default recommendation for new builds, with low, medium, and high quality tiers, and notes that low is especially strong for latency-sensitive use cases.

That split matters because OpenAI is pitching the model for production surfaces, not just showcase prompts. Cheap low-end renders and more expensive high-fidelity passes fit that story better than a single headline number.

A separate HN thread from June pushed the story past launch-week novelty. As the Ask HN summary notes, commenters described image models crossing into replace-a-real-step territory, especially for e-commerce product photography and infographic-style visuals.

The interesting part is how narrow those claims were. People were not claiming image models had solved all design work. They were naming specific asset classes where the output was suddenly usable enough to change the workflow.

The fastest way people checked the upgrade was by throwing structured prompts at it. The HN discussion roundup captured Simon Willison's "where's the raccoon holding a ham radio" test, a crowded-scene prompt that forces the model to manage object placement and detail density at the same time.

That same roundup also pointed to minimaxir's benchmark prompt, an 8 by 8 grid of Pokémon whose National Pokédex numbers match the first 64 prime numbers. That is a good example of where the release stands out: the conversation around this model kept drifting toward grids, labels, panels, and layout rules, which is exactly the kind of work older image demos usually avoided.