GPT Image 1.5 Review: How Good Is OpenAI's New Image Model?

I've been using ChatGPT for image generation since it launched, and honestly? It was frustrating. Every time I tried to edit something simple, like changing a shirt color, the face would morph into someone completely different. I'd spend 20 minutes regenerating the same image trying to get basic edits to stick.

Then OpenAI dropped GPT Image 1.5 on December 16, 2025. I've spent the last few days putting it through its paces, and here's the thing: it actually fixes most of what annoyed me about the original.

Quick Answer: GPT Image 1.5 is a legitimate upgrade. The editing preservation is dramatically better, text rendering actually works now, and the 4x speed improvement is real. It's not perfect, and I'll get into the limitations, but for the first time ChatGPT feels usable for actual creative work rather than just playing around.

What Actually Changed?

Look, I'm not going to give you the marketing speak. Here's what matters.

OpenAI rebuilt how the model handles edits. The old version would essentially regenerate the entire image when you asked for changes. The new version actually understands what you want to keep versus what you want to change. Sounds obvious, but this was broken for months.

The text rendering improvements come from better training data and architecture changes. I don't know the technical details, but the results speak for themselves. I generated a fake newspaper layout yesterday and could actually read the articles. On the old model, that would have been gibberish.

Speed improvements come from infrastructure optimization. Nothing groundbreaking technically, just smart engineering.

The Editing Is Actually Good Now

I want to show you what I mean about the editing improvements because this is the biggest deal.

The Old Problem

Here's what would happen before. You'd generate an image of a person, love their face, but want to change their outfit. You'd type "change the blue shirt to red" and get back... a completely different person wearing a red shirt. Same prompt, same seed, different human being.

I wasted so many hours on this. Generate, regenerate, cry, regenerate again.

What's Different Now

GPT Image 1.5 actually tracks what you're asking to change. In OpenAI's demo, they changed one person in a photo to anime style while keeping another person photorealistic. The model understood "keep the man on the right the way he is" meant preserve EVERYTHING about him, not just his position.

I tested this myself with a headshot I'd generated. Asked to change the background from an office to a beach. Same person, same lighting on their face, just a new background. This would have been impossible three months ago.

Adding and Removing Elements

The removal tool is legitimately impressive. I had an image with an object I didn't want, asked to remove it, and the model filled in the space intelligently. No obvious smearing or artifacts.

Adding elements works similarly well. The new stuff matches the existing lighting and perspective rather than looking pasted in.

Text Rendering Finally Works

Real talk: text rendering on AI images has been a joke. Every model produces beautiful images with nonsense where the text should be. "COFEFE SHOP" instead of "COFFEE SHOP." Random characters that look like letters but aren't.

GPT Image 1.5 is the first model I've used where I'd actually trust it for graphics that include text.

What I Tested

I generated a fake magazine cover with a headline, subheadings, and body text. Probably 200 words total. Results? Maybe 90-95% accuracy. A few typos in the fine print, but the headlines were perfect.

For social media graphics, blog headers, mockups... this is good enough. You might need to regenerate once or twice for perfection, but you're not fighting the model anymore.

The Limitations

Multilingual text is still rough. I tried generating text in Japanese and got mixed results. If you need accurate non-English text, you're still better off adding it in post-production or using specialized tools.

Very small text also struggles. Labels on products, fine print, anything below maybe 12pt equivalent. It's better than before, but not reliable.

The Speed Difference Changes Everything

I know "4x faster" sounds like marketing fluff. It's not. This genuinely changes how you work with the tool.

Before: Every Generation Was Precious

When each image took 20-30 seconds, you'd craft your prompt carefully. Accept "good enough" results. Avoid experimentation because failed attempts cost real time.

Now: Iterate Like You Mean It

At 5-8 seconds per generation, you can actually explore. Try something, see what happens, adjust, try again. I generated probably 50 images yesterday experimenting with different approaches. That would have taken over an hour before. Now it's 15 minutes.

This is how AI image generation should feel. Quick enough that you're not losing creative momentum waiting for results.

Comparison: Where Does It Stand?

Here's my honest assessment compared to other tools I use.

Versus Midjourney

Hot take: for practical work, GPT Image 1.5 is now more useful than Midjourney for me. Midjourney still produces that distinctive aesthetic some people love, but the editing capabilities in GPT Image 1.5 mean I can actually iterate on results rather than regenerating from scratch.

Your mileage may vary. If you love the Midjourney look, keep using it.

Versus Local Models

If you're running ComfyUI or similar, you still have more control. LoRAs, ControlNet, custom workflows. But for quick generation without setup? GPT Image 1.5 is compelling.

For projects where I need maximum consistency or specific styles, I still use local tools or Apatero.com for the workflow flexibility. For quick concepts and iterations, ChatGPT is now genuinely useful.

What Still Sucks

I promised honest assessment, so here's what's still broken.

Character Consistency Across Sessions

You cannot reliably reproduce the same character in a new chat. Generate a person you love, come back tomorrow, and you'll get someone different even with the same prompt.

For projects requiring persistent characters, this is a dealbreaker. You need dedicated tools, LoRA training, or platforms like Apatero.com that handle consistency properly.

I've written about creating consistent AI faces if you need solutions for this.

Multiple Faces in One Image

Crowd scenes still look weird. The faces are individually fine but collectively odd. Too similar, wrong diversity distribution, uncanny valley territory.

OpenAI's Content Restrictions

Some legitimate creative work gets blocked by safety filters. I'm not talking about anything sketchy, just normal creative requests that happen to trigger false positives. This is frustrating when you're in a creative flow.

No Offline Option

Everything runs through OpenAI's servers. If you need to work offline, need complete privacy, or want to avoid recurring costs at scale, this isn't your tool.

Who Should Actually Use This?

Great For:

• Quick concept art and ideation
• Marketing graphics and social media content
• Blog illustrations (like this one)
• Product mockups and variations
• Anyone who wants AI images without technical setup

Maybe Not For:

• Projects requiring consistent characters across sessions
• High-volume production where API costs matter
• Work requiring complete creative control
• Anything that might trigger content filters

The Sweet Spot

If you're already paying for ChatGPT and occasionally need images, this is now actually useful rather than a frustrating toy. The upgrade happened automatically. Just start asking for images.

If you're building something more serious with AI images, treat GPT Image 1.5 as a prototyping tool. Use it for quick exploration, then move to dedicated tools for production.

Frequently Asked Questions

Is GPT Image 1.5 free?

Available to all ChatGPT users including free tier. Rate limits are lower for free users. Plus/Business/Enterprise get more generations.

How do I access it in the API?

Use model identifier gpt-image-1.5. Same endpoint structure as before. Check the OpenAI Playground to test.

Can it generate NSFW content?

No. OpenAI's content restrictions remain. For unrestricted generation, you need alternative platforms.

How does it compare to DALL-E 3?

It effectively replaces DALL-E 3 for most uses. Better editing, faster generation, better text. No reason to use the older model unless you specifically prefer its behavior.

Can I use the images commercially?

Yes, per OpenAI's terms. Standard usage rights apply.

What resolution does it output?

Fixed resolutions. For higher resolution, you'll need upscaling tools afterward.

Does it work on mobile?

Yes. Same capabilities across ChatGPT apps and web.

What happened to the old model?

Still available as a custom GPT for workflows built around it.

Is the 4x speed real?

In my testing, yes. 20-30 seconds dropped to 5-8 seconds for typical generations.

Can I edit just part of an image?

Yes. Describe what to change and what to keep. The model handles it intelligently.

The Bottom Line

GPT Image 1.5 is the first ChatGPT image update that feels like genuine progress rather than incremental improvement. The editing preservation alone makes it worthwhile. The text rendering makes it practical for real work. The speed makes it enjoyable to use.

Is it going to replace dedicated tools? No. For serious production work, character consistency, or specific creative requirements, you still need proper solutions.

But for quick ideation, prototyping, and casual creation? It's actually good now. I never thought I'd say that about ChatGPT images.

Give it a try. Generate something, ask for an edit, see if the face stays the same. You might be surprised.