ChatGPT Images 2.0 ships gpt-image-2 with stronger text layout control

• OpenAI shipped ChatGPT Images 2.0 in ChatGPT, while the main HN launch thread links to the same-day OpenAI announcement and the API-facing gpt-image-2 rollout.
• OpenAI is pitching the upgrade around structured visual work, and the HN core summary centers the gains on better text rendering and layout control for comics, infographics, and other production-style assets.
• The API side is not a side note: according to the HN discussion summary, developers got a new model card, mostly unchanged pricing versus the prior image model, and slightly lower output cost, which lines up with OpenAI's developer announcement.
• Community reaction split along a familiar line, because the HN discussion summary highlights prompt-adherence praise while the HN core summary also surfaces claims that Nano Banana still wins on raw image fidelity.
• The most useful creative-side read came later, when the Ask HN core summary captured product-photo and infographic workflows that users said had crossed from demo material into commercially usable output.

You can read the launch post, skim the new gpt-image-2 model page, and dig into OpenAI's own prompting guide. For an immediate side-by-side stress test, Simon Willison's raccoon prompt is more revealing than most announcement copy.

OpenAI shipped ChatGPT Images 2.0 on April 21 as both a ChatGPT product update and an API model rollout. The official package was simple: a new image system in ChatGPT, plus gpt-image-2 for developers via the API and Codex, per the OpenAI announcement and developer announcement.

The important part is where OpenAI aimed it. In OpenAI's launch framing, this was about complex, polished visuals that are ready to use, not just prettier concept art.

OpenAI's strongest claim was not photorealism. It was control over text and structure.

According to the HN core summary, the upgrade was framed around comics, infographics, and other layouts where image models usually fall apart once typography and composition have to survive the same prompt. OpenAI's prompting guide points at the same lane and explicitly calls out complex structured visuals, including infographics, diagrams, and multi-element marketing assets.

That guide also makes the workflow more concrete:

• gpt-image-2 is the recommended default for new builds.
• outputQuality supports low, medium, and high.
• OpenAI says low is strong enough for latency-sensitive work.
• It suggests stepping up to medium or high when fidelity matters most, especially for dense text and infographic-style output.

This is the part creative teams actually care about. Clean type and stable layout are what move an image model from moodboard toy to production helper.

The API model page describes gpt-image-2 as OpenAI's state-of-the-art image generation and editing model, with support for text and image input, image edits, and flexible image sizes through the model docs.

OpenAI's own guide adds the constraints that matter in real pipelines:

• both edges must be multiples of 16
• the maximum edge length is 3840 px
• aspect ratio can go up to 3:1
• total pixels must stay within the model's allowed range

Pricing was not marketed as a major change, and the HN discussion summary reflects that. In OpenAI's developer announcement, the listed rates were $5 per 1M text input tokens, $10 text output, $8 image input, and $30 image output, with cached-input discounts on both modalities.

That combination, stable pricing plus more explicit layout-oriented guidance, makes the release feel less like a speculative model jump and more like OpenAI trying to normalize image generation as a standard app surface.

The first wave of discussion was not really about whether the model had improved. It was about which kind of improvement mattered.

According to the HN discussion summary, one early HN thread focused on prompt adherence and pricing, while the HN core summary captured the opposing complaint that raw image fidelity still lagged Nano Banana for some users. That split shows up in Simon Willison's comparison, where gpt-image-2 looked meaningfully stronger than gpt-image-1 on a busy "Where's Waldo" style prompt, but still benefited from quality tuning and iteration.

The practical read is straightforward: OpenAI appears to have moved the needle fastest on images that need instructions, labels, and composition to stay intact. The fidelity crown was still being argued in public.

The clearest commercial-use evidence in this source set came from a different HN thread a few weeks later. There, the Ask HN core summary preserved a creative-side report that Nano Banana Pro could turn a bad iPhone photo into usable whitebox product shots and Amazon-style infographic images.

the Ask HN discussion summary groups that example with a broader pattern of users describing faster idea-to-production workflows, but the product-image example is the one worth bookmarking. It names the breakpoint clearly:

• rough phone photo in
• usable product photo out
• infographic-style marketplace assets in the same flow
• work that the commenter said would previously have needed a photographer and a designer

That does not prove ChatGPT Images 2.0 wins every head-to-head on beauty. It does show where this whole category crossed an economic threshold: structured commerce visuals, not gallery pieces, are where the newest image models started looking less like experiments and more like replacement-grade tooling.