GPT Image 2 Review: What Replaced DALL-E 3 Inside ChatGPT

If you’ve generated an image inside ChatGPT recently, you probably weren’t using DALL-E 3 — even if that’s still the name most people search for. OpenAI fully retired DALL-E 3 in May 2026, closing out both the API and the last legacy integrations, and replaced it with GPT Image 2, part of what OpenAI calls ChatGPT Images 2.0. We spent a week putting the new model through its paces to see what actually changed, and whether the bundled-with-ChatGPT pricing model still makes it the easiest AI image generator to recommend to non-designers.

Quick Take: GPT Image 2 is faster, sharper, and far better at legible text than DALL-E 3 ever was, and it’s bundled into the same ChatGPT Plus subscription millions of people already pay for. It won’t out-art Midjourney and it isn’t trying to, but as an accessible, conversational, all-in-one image tool, it’s currently the best free-with-your-subscription option available.

Why DALL-E 3 Is Gone — and What It Became

DALL-E 3’s replacement inside ChatGPT actually began quietly back in 2025, with OpenAI shifting image generation into a model that understands images and text within the same underlying network, rather than routing prompts out to a separate specialist model. GPT Image 2, launched in April 2026, is the latest and most capable version of that architecture. The upgrade shows immediately: photorealism convincingness scores jumped substantially in blind testing, prompt adherence is noticeably tighter on complex multi-part instructions, and — most importantly for everyday users — text inside images finally looks like actual words instead of a rough approximation.

The trade-off is generation speed. Where DALL-E 3 could return an image in ten to fifteen seconds, this newer, more capable architecture takes considerably longer per image, which is a real friction point if your workflow depends on rapid iteration across dozens of variations in a single sitting.

Hands-On Testing

We tested GPT Image 2 through its “Thinking” mode, available to paid subscribers, which lets the model reason through a request — including pulling in live web context mid-generation — before producing up to several coherent images from a single prompt. For anyone using ChatGPT conversationally (describe what you want in plain language, get an image, ask for a tweak, get an updated version), this is a genuinely more natural workflow than the prompt-engineering culture that surrounds tools like Midjourney.

Text rendering was the standout improvement in our testing. A poster prompt with a full headline and a subhead came back legible and correctly spelled on the first attempt in the large majority of trials — a dramatic jump from DALL-E 3’s inconsistent record on the same category of prompt. Photorealistic portraits also looked noticeably more convincing, with fewer of the plasticky skin textures and uncanny eye rendering that dogged earlier OpenAI image models.

Native resolution has also climbed meaningfully, with high-resolution output now standard rather than a paid add-on, and higher resolutions available in beta for users who need larger source files for print or large-format use.

The ChatGPT Integration Advantage

The single biggest reason to consider GPT Image 2 over a dedicated tool isn’t raw image quality — it’s the fact that you’re already inside a conversation. Instead of crafting a precise, jargon-heavy prompt from scratch, you describe roughly what you want, and the underlying model expands and refines that description before generating. Because generation happens inside the same model handling your text conversation, GPT Image 2 can reference earlier messages, understand relationships between multiple images in a thread, and make targeted edits without you having to re-describe the entire scene from zero.

Commercial usage rights are included by default, even on the free tier, which is a meaningfully more generous starting point than most competitors offer.

We also tested how well the model handled a multi-image sequence — asking for a set of four consistent illustrations for a short explainer, referencing the first image’s style when requesting the second, third, and fourth. Consistency held up reasonably well across the set, though not quite as reliably as Ideogram’s dedicated character-reference tooling or Midjourney’s Omni Reference feature, both of which are purpose-built specifically for that use case.

Pricing Breakdown

For light-to-moderate personal use, the practical math is simple: if you already pay for ChatGPT Plus for its text and reasoning features, image generation now effectively costs nothing extra on top.

Pros and Cons

What We Liked

• Bundled into the same $20/month ChatGPT Plus subscription millions already pay for
• Text rendering is dramatically better than DALL-E 3 and competitive with Ideogram
• Conversational, iterative editing workflow is more approachable than prompt-engineering-heavy tools
• Commercial usage rights included by default, even on the free tier
• Content provenance metadata included on every generated image

What Held It Back

• Generation is noticeably slower than the DALL-E 3 model it replaced
• Artistic range and painterly style options trail Midjourney
• Rate limits inside ChatGPT Plus are unpublished and appear to fluctuate
• DALL-E 3 API is fully retired, so any old integrations must migrate immediately

GPT Image 2 vs. the Competition

Against Midjourney, GPT Image 2 wins decisively on accessibility and text rendering but loses on pure artistic polish. Against Ideogram, the text-rendering gap that used to favor Ideogram alone has narrowed considerably, though Ideogram’s brand-consistency tooling remains more purpose-built for ongoing campaign work. Against Adobe Firefly, GPT Image 2 offers a more natural, conversational interface, while Firefly counters with deeper design-software integration and commercially licensed training data.

Migrating Away From the Old DALL-E API

Anyone with existing code still pointed at the old DALL-E 3 endpoint needs to move immediately — the API was fully shut down in May 2026 and now returns errors on every call. The migration itself is straightforward in most cases: swapping the old model identifier for the current GPT Image model identifier, or using OpenAI’s auto-updating alias so future model upgrades apply without further code changes. Teams that built tooling around DALL-E 3’s specific response format should budget a little extra testing time, since the newer model’s output structure and available parameters differ in a few small but meaningful ways.

For developers billing by usage, it’s worth re-checking per-image API costs after migrating — pricing on the newer model has shifted since launch, and in our testing the effective cost for high-quality, large images had actually drifted downward slightly compared to the ceiling most guides originally quoted, narrowing the gap with older, cheaper models.

Tips for Better Results Inside ChatGPT

Because GPT Image 2 benefits from conversational back-and-forth rather than a single, perfectly engineered prompt, we found the best results came from starting broad and refining iteratively: describe the general scene first, review the output, then ask for specific, targeted changes (“make the lighting warmer,” “move the subject to the left third of the frame”) rather than rewriting the entire prompt from scratch each time. This plays directly to the model’s strength of referencing earlier context in the same thread, and it avoids losing elements of the composition you’d already gotten right.

For text-heavy requests specifically, being explicit about the exact wording, in quotation marks, produced meaningfully more accurate results than describing the text indirectly — the same technique that helps on dedicated text-rendering tools like Ideogram.

Frequently Asked Questions

Do free ChatGPT users get GPT Image 2? Yes, but through a faster, more limited “Instant” mode rather than the full Thinking mode reserved for paid subscribers, and with a lower daily generation cap.

Is DALL-E 3 still available anywhere? No — both the consumer-facing integration and the developer API were fully retired in May 2026. Any reference to “using DALL-E” inside ChatGPT today actually refers to GPT Image 2.

Can I use these images commercially? Yes, commercial usage rights are included by default across both free and paid tiers, which is more generous than several competitors that gate commercial rights behind a paid plan.

Who Should Use GPT Image 2?

Existing ChatGPT Plus subscribers who need occasional, high-quality images without learning a second tool are the clearest fit. Non-designers who find prompt engineering intimidating will especially appreciate the conversational back-and-forth. Power users doing high-volume, rapid-iteration work may find the slower generation time a genuine bottleneck compared to more specialized alternatives.

How We Tested

We ran our standard fifteen-prompt baseline test set through GPT Image 2 inside ChatGPT Plus over one week, alongside a dedicated set of text-rendering prompts (posters, headlines, and labeled diagrams) to directly compare accuracy against Ideogram and Midjourney on the same requests. We also timed generation speed across both Instant and Thinking modes to quantify the trade-off between the two, and reviewed the API documentation and current per-image pricing directly against OpenAI’s published rate card to confirm costs hadn’t drifted since the model’s April 2026 launch.

Final Verdict

Rating: 4.4/5. GPT Image 2 is a substantial upgrade over DALL-E 3 in almost every measurable way, and bundling it into ChatGPT Plus remains one of the smartest pricing decisions in the category. It’s not the tool for maximum artistic ambition, but as an accessible, all-purpose generator that just about anyone can pick up immediately, it’s hard to beat.