ChatGPT Images 2.0 adds multilingual text and 8-image consistency in creator tests

• OpenAI's official launch post says ChatGPT Images 2.0 improves instruction following, text rendering, and structured visuals, while the linked launch summary adds multilingual support and a new eight-image consistency feature.
• For creators, the sharpest upgrade is in layouts that used to break image models: posters, diagrams, comics, slides, and other text-heavy compositions, according to the main HN thread and the OpenAI developer announcement.
• The rollout also reaches production surfaces, not just the ChatGPT UI: the API model page positions gpt-image-2 as OpenAI's state-of-the-art generation and editing model, and the discussion roundup points readers to pricing and builder tests.
• Early HN testing focused on prompt adherence and heuristic-heavy prompts, where one builder test and another experiment treated the model less like a vibe machine and more like a layout engine.
• The same HN thread that praised the image quality also surfaced a harder production question: one top comment argued the outputs are becoming hard to distinguish from non-AI work, while the thread summary notes ongoing style-credit concerns.

You can read OpenAI's launch post, jump straight to the API model page, and check the image generation guide for the production workflow. The HN thread also links directly to a pricing discussion, a prompt-adherence test, and a heuristics-heavy experiment.

The headline improvement is not raw prettiness. It is control over structured visuals.

According to the launch summary, ChatGPT Images 2.0 targets dense text, multilingual rendering, complex compositions, and aspect-ratio flexibility. The developer announcement makes that even more concrete by naming diagrams, infographics, charts, posters, and comics as first-class outputs.

For creative teams, that list matters because it covers the formats that usually force a cleanup pass in Figma or Photoshop. OpenAI is explicitly pitching higher-fidelity text-heavy visuals, not just prettier standalone images.

OpenAI's launch post and the HN-linked summary both point to a new "thinking" layer that can reason before generating. In the rollout description, that feature can also use web-sourced information and generate up to eight consistent images from a single prompt.

That is the more interesting creator unlock than a single hero image. A model that can keep a character, layout system, or visual premise stable across eight frames is much closer to storyboard work, comic pages, campaign variations, and product sequence mockups.

The production path is already in place. The image generation guide says gpt-image-2 supports both generations and edits, plus multi-turn editing through the Responses API, while the API model page positions snapshots as a way to lock behavior to a specific version.

The best early community signal came from people immediately stress-testing prompt adherence instead of posting glamour shots.

HN commenter Simon Willison's prompt-adherence test used a "Where's Waldo"-style scene to see whether the model could hold onto a very specific search task. Another HN experiment from minimaxir's heuristics-heavy prompt pushed on domain knowledge and tool use rather than surface aesthetics.

That matches the rest of the rollout. The developer announcement frames gpt-image-2 as useful for apps, ads, social, presentations, docs, and product flows, with resolutions up to 2K. The model is also available in the API and Codex on day one, and the same announcement lists pricing at $8 per 1M image input tokens, $2 cached image input tokens, $30 image output tokens, and $5 per 1M text input tokens.

The strongest negative reaction in the thread was not about quality. It was about believability.

As one top HN comment put it, the images are now good enough that they can be hard to identify as AI. That same comment also raised the older complaint that generative systems learn specific artistic styles without clear credit to those artists, and the thread summary flags both authenticity and style appropriation as live concerns around the launch.

Those concerns sit awkwardly next to the product pitch, because the very features OpenAI is celebrating, better text, better layout, better consistency, are the ones that make generated work easier to ship as finished design.