ChatGPT Images 2.0 and gpt-image-2: OpenAI’s new flagship, API sizes, and how it fits the stack

In April 2026, OpenAI announced ChatGPT Images 2.0 alongside the API model GPT Image 2 (gpt-image-2). The model page describes a state-of-the-art text- and image-to-image stack and points to the image generation guide, pricing, and cost calculators.

This post is a map of first-party sources, not a hands-on review. Product marketing (social posts, “thinking,” leaderboards) moves fast; treat benchmark headlines as pointers to re-check, not as specs.

What builders should anchor on

1. Model name: OpenAI documents gpt-image-2 and snapshot gpt-image-2-2026-04-21 on the model index.
2. Surface area: The image generation guide covers the Image API (generate / edit) and Responses API (conversation + image_generation tool), with guidance on which to pick.
3. Resolutions and quality — the guide lists common sizes (e.g. 1024×1024, 1536×1024, 1024×1536, 2048×2048, 2K landscape, 4K-class) with pixel and aspect constraints, and quality: low | medium | high | auto. It notes that 2K+-class outputs can be experimental, and that gpt-image-2 does not support transparent backgrounds in the current guide.
4. Limitations (docs) — latency (complex prompts can be long), text placement, consistency for brands/characters, layout precision; see the Limitations section.
5. Access — org verification may be required for GPT Image models on the API; see the guide.

How this connects to the rest of our blog

• How diffusion works (generic): How do image generation models work? — denoising, VAE latents, CFG, and a noise-to-image strip at /blog/diffusion/noise-to-image.png.
• “Images in chat” is a different product shape than a plain LLM context window; tokens matter when a chat model orchestrates tool calls, while per-image costs follow OpenAI’s image generation pricing and calculator in the same guide.

Read next

• OpenAI — Introducing ChatGPT Images 2.0
• OpenAI — Image generation API guide
• OpenAI — GPT Image 2 model
• How diffusion image generation works (ExplainX)

Sizes, quality labels, and pricing are versioned. Re-check OpenAI’s platform docs and your plan before building dependencies.