ChatGPT Images 2.0 compares with Nano Banana on hard prompts

• OpenAI says the launch summary improves text rendering, instruction following, and consistency for infographics, comics, and other layout-heavy images, while the HN discussion highlights show early testers immediately stress-testing those claims on hard prompts.
• The first HN verdict was mixed: the discussion highlights quote neom saying their prompt set was still nowhere near Nano Banana, but the core HN summary also notes edge cases where gpt-image-2 beat Nano Banana.
• OpenAI tied the release to a new "thinking" mode that can use reasoning and web search, and the launch summary says it can generate up to eight consistent images from one prompt.
• For creators, the interesting gap is between structured design work and adversarial weird-prompt work: the HN core summary frames comics and text-heavy layouts as the main upgrade, while HN commenters kept finding failure cases on stranger prompts.

You can read OpenAI's launch post, the API model page, and the developer announcement. The HN thread is where the useful mess starts, including Simon Willison's hands-on tests and a quick round of Nano Banana comparisons. OpenAI also published a system card and a prompting guide, which is where the workflow knobs show up.

OpenAI's pitch is straightforward. ChatGPT Images 2.0, powered by gpt-image-2, is supposed to be better at multilingual text, tighter instruction following, and image sets that stay visually consistent across multiple outputs.

The official materials point to the same creative use cases over and over:

• infographics
• floor plans
• charts and posters
• comics and manga-style panels
• slide-like or doc-ready visuals
• image series with up to eight consistent outputs from one prompt

The developer announcement adds two concrete production details missing from most reaction posts: support for more aspect ratios and resolutions up to 2K, plus positioning around assets that need to be readable, localized, and usable without much cleanup.

The first day community read was less victory-lap, more prompt cage match. In the HN thread, minimaxir went straight to the model card and pricing, Simon Willison started running oddball prompts through the API, and other commenters compared results against Nano Banana almost immediately.

Three patterns surfaced fast:

• According to the HN discussion highlights, neom ran prompts on both ChatGPT and the API and said none of them were close to Nano Banana on their test set.
• According to the core HN summary, justani reported a split result: some prompts still failed in both systems, but at least one case that Nano Banana missed worked in gpt-image-2.
• In the full HN thread, commenters kept using comics, menus, and weird object-scene prompts because those expose whether better typography also comes with better compositional logic.

That is a more useful first-day frame than the polished demos. The new model looks aimed at real design tasks, but HN testers treated it like a benchmark for ugly edge cases.

The sharpest early takeaway is that the upgrade seems strongest where the image needs structure. The HN core summary specifically calls out comics, infographics, and other multi-panel or layout-heavy formats as the main beneficiary of the new text rendering and consistency work.

But the HN thread also produced a cleaner failure description than the launch copy did. One commenter in the discussion argued that Nano Banana often gets the logic and punts on the art, while gpt-image-2 gets the art and punts on the logic. That is not a benchmark result, but it does fit the mixed reports in the HN discussion highlights and the core summary.

For creative workflows, that split matters more than leaderboard language. A model that can hold panel layouts, labels, and repeated characters together is useful in one class of jobs. A model that still breaks on bizarre constraint-heavy prompts has not closed the whole gap.

The last useful detail is that this was not just a ChatGPT front-end refresh. OpenAI's developer announcement shipped gpt-image-2 into the API and Codex on day one, with posted token pricing and support for both generations and edits.

The prompting guide is also unusually specific about workflow tradeoffs. It recommends gpt-image-2 as the default for new builds, says low, medium, and high outputQuality are all supported, and notes that low is especially strong for latency-sensitive use cases. That gives creators a more practical picture than the launch post alone: OpenAI is selling this as a production image model, not just a prettier demo model.