OpenAI launched ChatGPT Images 2.0 on Tuesday, April 21, 2026, its first image-generation model with reasoning built in. The update rolls out immediately to every ChatGPT and Codex user — free tier included — and ships alongside a developer API under the identifier gpt-image-2.

The headline capability is not resolution or speed. It is that the model can now search the web, plan a layout, and double-check its own output before committing a single pixel. That, plus a genuine step-change in text rendering, moves image generation from a creative toy into something OpenAI is pitching as a production-grade design tool.

An Image Model That Reasons

ChatGPT Images 2.0 runs in two modes. The default Instant mode behaves like a traditional diffusion model — one prompt, one image, fast. The new Thinking mode, available on Plus and Pro plans, routes requests through OpenAI's o-series reasoning stack before the image is rendered, according to OpenAI's launch materials as summarized by iClarified.

In Thinking mode, the model can search the web for real-time reference, transform uploaded documents into visual explainers, and analyze the structure of its own output before delivering it. TechCrunch noted that OpenAI declined in press briefings to say which underlying architecture powers the model — only that its reasoning capabilities allow it to double-check its creations, make multiple images from a single prompt, and pull in outside information.

When Thinking is on, a single request can return up to eight consistent images with character and object continuity — a multi-panel comic, a set of social media variants, a room-by-room interior layout — all from one prompt. PetaPixel highlighted OpenAI's framing: "Instead of prompting one image at a time and stitching the project together yourself, you can ask for a coherent set of up to eight outputs in one go."

Text Rendering, Finally

Every image model released in the last three years has a reputation for butchering text. Letters melt, words misspell, and anything more than a short caption ends up looking like a ransom note. OpenAI's pitch with Images 2.0 is that this specific failure mode is largely fixed.

The model is now able to render "small text, iconography, UI elements, dense compositions, and subtle stylistic constraints" at up to 2K resolution, OpenAI stated in the launch materials quoted by PetaPixel. The practical effect: readable typography in posters, legible axis labels in infographics, actual working interface mockups — the categories of work that previously required a designer to clean up every AI output by hand.

The gains extend well beyond English. OpenAI is calling out "significant progress" on non-Latin scripts, specifically Japanese, Korean, Chinese, Hindi, and Bengali, according to Engadget's coverage of the launch. That removes one of the most brittle parts of multilingual creative workflows — a model that can produce readable Hangul, Devanagari, or kanji inside a poster or diagram without the usual glyph drift.

Free for Everyone on Day One

The distribution choice is unusual. Rather than gating the core quality jump behind a premium subscription, OpenAI is shipping the base model to every ChatGPT tier on day one — Free, Plus, Pro, Business, and Codex users all get Images 2.0 as of Tuesday, The Verge reported.

Thinking mode is what separates the paid tiers. It is reserved for Plus, Pro, and Business subscribers, with Pro users getting an additional layer the company is calling ImageGen Pro on top, per the coverage in iClarified. In other words, the quality floor goes up for everyone, and the reasoning, multi-image, and web-search ceiling sits at the paid tiers.

Aspect ratio flexibility is part of the same bet on real-world use. The model now supports ratios from 3:1 ultra-wide down to 1:3 ultra-tall, as confirmed in ZDNET's early hands-on. That covers social banners, vertical mobile, editorial spreads, and print-oriented layouts without a post-processing crop step.

For Developers: gpt-image-2

The launch is not only a ChatGPT feature. OpenAI is exposing the same image stack to developers through the Image API under the model ID gpt-image-2, with outputs up to 2K resolution, iClarified noted. For conversational and multi-step workflows, the model is also available as a built-in tool in the Responses API.

There is a Codex tie-in as well. For Mac developers using Codex's background computer-control features, the image model is integrated directly into the workspace. A developer can generate multiple UI directions and prototypes, compare results, and turn the strongest ideas into live products without leaving the app, per OpenAI's launch materials.

Surface What It Is Where to Access
ChatGPT product Consumer image generation in chat threads (Instant and Thinking) Web and iOS/Android, all tiers
Image API One-shot generation and editing endpoint gpt-image-2
Responses API Conversational multi-step image workflows as a built-in tool gpt-image-2
Codex integration Native image generation inside the Codex Mac app workspace Codex app for Mac

Pricing varies by output quality and resolution; OpenAI has not yet published a simple flat-rate table, per PetaPixel. The model's knowledge cutoff is December 2025, TechCrunch noted, which matters for prompts that involve recent real-world references.

What It Doesn't Do

OpenAI has been unusually explicit about the limitations. The company notes that the model still struggles with tasks that require precise understanding of the physical world — origami folding and certain puzzle layouts are called out specifically. Extremely dense or repetitive textures can be challenging, and detailed technical diagrams may still require manual review, iClarified reported.

Complex prompts in Thinking mode can take up to two minutes to process — a practical consideration for anyone building this into a user-facing product, TechCrunch noted. And text rendering, while significantly improved, is characterized by OpenAI as "better, not solved." For typographically demanding work where a single character matters, a human design review pass is still the safe default.

Image generation as a product category has been waiting for the mid-tier competence problem to be solved: not the best-case showpiece, but the middle of the distribution — the posters, social assets, UI mockups, and educational diagrams that agencies and small teams actually need to ship every day. Images 2.0 is OpenAI's clearest attempt yet to aim at that middle. Whether the reasoning layer is enough to hold the lead over Google's Imagen line, Midjourney, and Flux through the next six months is the question the market now gets to answer.