GPT Image 2

Every image model before this generation shared one embarrassing weakness, DALL·E 3 included. Ask for a poster, a product label, a chart or a shop front, and the words came back as convincing gibberish — letter-shaped marks arranged like text by something that had never read anything.

That single failure kept generated images out of most commercial work. A beautiful mock-up you cannot put a real product name on is a mood board, not an asset.

GPT Image 2, released on 21 April 2026 under the product name ChatGPT Images 2.0, is the point at which OpenAI's answer to that problem became genuinely usable. It also changes how the model works: rather than generating in one pass, it plans the structure of the image first. This review covers what that buys, the documented constraints most write-ups skip, and where the model still loses to its rivals.

What Shipped On 21 April 2026

• Text rendering is the reason to care — dense paragraphs, small lettering, multilingual characters and complex layouts, which is what made previous image models unusable for anything with words in it.
• It reasons before it renders, planning image structure rather than generating in a single pass, which is why layouts hold together.
• Resolution goes to 3840px on the long edge, with both edges multiples of 16 and an aspect ratio no wider than 3:1.
• Editing supports masks and multiple reference images, and every input is processed at high fidelity automatically rather than being a setting you tune.
• Two API surfaces: the Image API for generation and edits, and the Responses API for images produced inside a conversation or an agent flow.

GPT Image 2: The Model Card In Plain English

Text Rendering Is The Headline Improvement

OpenAI's documentation is direct about the priority: stronger text rendering, layout fidelity, editing and multilingual support.

In practice this means three things that previous models could not do. Dense text — a paragraph rather than a word — comes out readable. Small lettering survives, which is what packaging, labels and chart axes require. Non-Latin scripts render as the characters they are meant to be rather than as decorative approximations.

The layout half is just as important and less discussed. An infographic is not text on an image; it is a spatial argument where the position of each element carries meaning. Producing one requires holding a structure in mind while rendering its parts, which is precisely what single-pass generation cannot do.

For anyone producing marketing assets, this is the difference between a model that generates a background for a designer and a model that produces something close to a finished comp. Product photography with accurate text on labels, logos and packaging is called out specifically in OpenAI's own materials, and that is a commercial use case rather than a demo.

The constraint on generated imagery was never beauty. It was that you could not put a real word in it.

Reasoning Applied To Pictures

The architectural change behind the layout improvement is that the model plans before it renders — researching and reasoning about image structure rather than producing the whole frame in one shot.

The practical consequence is that complex, multi-element prompts behave far better. Previous models degraded sharply as you added requirements: three specific objects in specific positions with specific text usually produced two of them, somewhere else, with the wrong words.

The trade-off is latency. Planning takes time, and a model that thinks about composition is slower than one that does not. OpenAI exposes quality tiers partly for this reason, and recommends square images and JPEG output where speed matters. For a batch of thumbnails, use the low tier; for a hero asset, do not.

The Resolution Rules Are Unusually Specific

Most image APIs offer a short list of sizes. GPT Image 2 accepts any resolution that satisfies a documented set of constraints, which is more flexible and requires more care.

The rules are: the longest edge may not exceed 3840 pixels, both edges must be multiples of 16, the ratio of long edge to short edge may not exceed 3:1, and the total pixel count must fall between 655,360 and 8,294,400.

Two of these catch people out. The multiple-of-16 requirement means an arbitrary size from a design spec will be rejected, so round before you request. The 3:1 limit means very wide banners — the kind a website header often needs — cannot be generated directly and must be produced at a legal ratio and cropped.

There is also a documented caveat worth respecting: outputs above roughly 2560×1440 are described as experimental. If you are building a pipeline that depends on 4K output, treat that as a feature under development rather than a guarantee, and keep a fallback path.

• Common safe sizes: 1024×1024, 1536×1024, 1024×1536, 2048×2048, 2048×1152.
• Large sizes available but flagged experimental: 3840×2160 and 2160×3840.
• Quality settings are low, medium, high and auto — low is genuinely useful for drafts and thumbnails.
• Output formats are PNG (default), JPEG and WebP, with a compression parameter on the latter two.
• Transparent backgrounds are in preview, and require PNG or WebP; JPEG cannot carry them.

Editing, Masks And The Four-Image Trick

The edits endpoint is where most production work actually happens, and it behaves differently from the selection tools designers are used to.

Masking here is prompt-based guidance. OpenAI's documentation states plainly that the model uses the mask as guidance but may not follow its exact shape with complete precision. That is an important expectation to set: this is not a lasso selection, and a pixel-exact edit against a hard boundary still belongs in an image editor.

Masks must match the input's format and size, stay under 50MB, and carry an alpha channel. What the mask buys you is a strong steer about where to work — enough for replacing a background, changing an object, or adjusting a region, and not enough for compositing against a precise edge.

The more interesting capability is supplying several reference images at once. OpenAI's own examples demonstrate four inputs in a single edit, which enables a genuinely useful pattern: give it the product, the brand palette, a style reference and a layout reference, and ask for a composition that respects all four.

One detail saves confusion. Unlike the previous generation, input fidelity is not a parameter you set — the model processes every input at high fidelity automatically. If you have code carrying that setting from an older integration, remove it.

Getting A Usable Asset Out Of It

The model rewards a different working method from its predecessors. Because it plans structure before rendering, front-loading the constraints produces far better results than iterating on a vague prompt.

This is the sequence that gets a production-ready asset with the fewest wasted renders.

1. Pin the snapshot in code — gpt-image-2-2026-04-21 — so an upstream model update cannot silently change output your brand depends on.
2. Draft at the low quality tier and a square 1024 by 1024. Composition problems show at that size, and you will find them for a fraction of the cost and latency.
3. Write the exact text you want rendered in quotation marks in the prompt, rather than describing it. The model can render words accurately, but only the ones you actually specify.
4. Round your target dimensions to multiples of 16 before requesting them, and check the long-to-short ratio is within 3:1 — a rejected request is the most common integration error.
5. Once composition is right, re-render at high quality and your final resolution, keeping the prompt identical so only the quality variable changes.
6. For edits, supply the product, palette, style and layout as separate reference images in one call, and use the mask as a steer rather than expecting it to be honoured exactly.

Where GPT Image 2 Is Weakest

A model this strong deserves an honest account of what it does not do, because the gaps are structural rather than temporary.

It is hosted only. There are no open weights, no self-hosting and no offline path. For a studio with confidentiality obligations over unreleased product imagery, or anyone in a jurisdiction with data residency requirements, that is a threshold question rather than an inconvenience — and it is where open-weight alternatives keep their advantage regardless of quality. Suite-tied models such as Adobe Firefly trade raw capability for provenance and integration instead.

Moderation is adjustable but not removable: the setting offers standard or less restrictive filtering, with no way to switch it off. For most work this never surfaces. For medical, forensic or certain artistic contexts, occasional refusals on legitimate requests are part of the deal.

Style is the other honest limitation. Models with a strong aesthetic signature — the reason many people still reach for Midjourney — produce a distinctive look that a general-purpose model tuned for accuracy and instruction-following does not. Precision and character are different goals, and this model chose precision.

Finally, provenance is not documented in the API reference. If your workflow depends on embedded content credentials being present in the returned file, verify that yourself rather than assuming it, and keep your own record of what was generated and when.

How OpenAI Charges For Images

Image generation is metered rather than subscribed, and the unit is worth understanding before building anything at volume.

Cost varies with what you ask for: the quality tier and the output resolution both move it, because both change how much work the model does. A low-quality 1024×1024 draft and a high-quality 3840×2160 render are not comparable line items. OpenAI publishes a calculator for exactly this reason.

Access is also tiered by rate limit rather than by feature. Usage tiers govern tokens per minute and images per minute, with the entry tier permitting only a handful of images a minute and the top tier permitting hundreds. For a batch pipeline this is the constraint that will bite first, well before spend does.

The batch endpoint is the lever most teams miss. Work that does not need to be interactive — overnight catalogue regeneration, bulk variants — belongs there rather than in the synchronous path.

Inside ChatGPT the model is bundled with a subscription instead, which is the better route for occasional use. We do not publish amounts here, since both the per-image rates and the plan structure change; read OpenAI's pricing page alongside the calculator.

Image Generation Terms Worth Knowing

Five terms from the documentation that determine what you can actually build.

Inpainting

Regenerating part of an existing image while leaving the rest intact. Here it is guided by a mask, but the mask is a steer rather than a precise boundary.

Input fidelity

How faithfully the model preserves detail from a supplied image. On this model it is fixed at high and is not a parameter you set — a change from the previous generation.

Partial images

Progressive renders streamed while generation completes, up to three. Useful for showing a user something is happening on a slow, high-quality render.

Quality tier

Low, medium, high or auto. Controls how much computation the model spends, which changes both latency and cost. Low is genuinely usable for drafts.

Snapshot

A pinned, dated model version — here gpt-image-2-2026-04-21. Pin it in production so an upstream update cannot silently change your output.