OpenAI dropped GPT Image 2 onto the API quietly. Same medium and high tiers as 1.5, same 6 credits per medium generation on Picturesque, same prompt parameters, same reference image cap at 16. On paper this looks like a minor revision. After running it through the same template library I have been using on Nano Banana Pro for the past several days, the actual gap between "minor revision" and "closing in on the frontier" turns out to be smaller than the version bump implies.
Here is the honest take. What stood out, what still bugs me, and where it has landed in our workflow.
The pricing setup
GPT Image 2 keeps the exact same pricing as 1.5 on Picturesque. 6 credits for a medium generation. high tier costs more (the credit jump matters, it is not a small bump). Medium is the daily-driver setting. high is for the hero shot you actually publish.
The price comparison that matters is against Nano Banana Pro. Nano Banana Pro sits one tier up at 12 credits and has been our default for character work, brand consistency, and any template that needed a model with serious headroom since Google released it. GPT Image 2 at half the price would only be interesting if the quality came along too.
It mostly does.
The quality gap is genuinely smaller now
I spent a long weekend running roughly 40 prompts through both models in parallel. Same prompts, same references where applicable, evaluated side by side at 100 percent zoom.
GPT Image 1.5 used to lose to Nano Banana Pro in a few specific places. Faces would drift slightly across a series. Brand colors would shift one or two shades when re-prompted. Hair detail at native resolution was a notch behind Google. It was still a great model, but you could tell where the budget had been spent.
GPT Image 2 has closed most of those gaps. Faces hold consistently across a 6 to 8 image character sheet. Color reproduction is meaningfully tighter than 1.5, even without explicit color references. Hair and fabric textures have more strand-level detail. For most prompts that are not specifically optimized for Nano Banana Pro's strengths, the output now sits in the same approximate quality tier.
That matters for our template library. We had been running a chunk of our character and portrait templates exclusively on Nano Banana Pro. As of this week, those templates also accept GPT Image 2 as an alternate model, because the quality difference is no longer big enough to justify forcing the 12 credit price on every generation. Users get to choose.
Where it is still off: the AI Slop look
Honesty time. GPT Image 2 has a tendency I do not love. It overpolishes.
There is a specific look that the OpenAI image stack has that you might call "AI Slop" if you are being unkind, or "the OpenAI house style" if you are being polite. Skin gets a subtle plastic sheen. Backgrounds get an extra pass of soft contrast. Highlights bloom slightly more than they would in a real photo. The whole image feels like it has been run through an Instagram filter that nobody asked for.
GPT Image 1.5 did this. GPT Image 2 still does it. Maybe a hair less than 1.5, but it is still there.
For stylized illustration or cinematic compositions you usually do not notice, because the polish reads as deliberate aesthetic choice. Where it bites you is realistic portraits, photojournalism style work, and any image that is supposed to feel candid or unretouched. The model just refuses to give you "raw." It always wants to make the image look like a magazine cover.
Workarounds help a little. Adding "unretouched, raw photo, slight grain, no airbrushing" to your prompt dampens the polish. Dropping from high to medium also helps because the model has fewer cycles to refine. You cannot fully turn it off though.
Nano Banana Pro does not have this problem to the same degree. It tends to produce cleaner, more neutral output that respects whatever you described instead of layering its own aesthetic on top. That is one of the genuine reasons we will keep Nano Banana Pro as the default for templates aimed at realism.
The aspect ratio cap is the real downgrade
This is the spec that frustrates me most about GPT Image 2. You get three aspect ratios: 1:1, 2:3, and 3:2.
That is it.
For comparison, Nano Banana Pro on Picturesque supports 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, and 21:9. Ten options covering essentially every modern delivery surface from square Instagram tile to ultra-wide cinematic poster.
Three aspect ratios means: square, vertical portrait, and horizontal landscape. That covers basic editorial work. It does not cover the modern social-first content stack:
- Instagram Reels and TikTok cover frames want 9:16. GPT Image 2 cannot natively produce 9:16. You generate at 2:3 and crop, losing headroom or footroom every time.
- YouTube thumbnails want 16:9. Generating at 3:2 and cropping loses more of the frame than you expect, especially if you composed for the 3:2 framing.
- Cinematic posters and ultra-wide hero images want 21:9. Forget it. Generate landscape, crop aggressively, or pick a different model.
- Pinterest wants 4:5 or 2:3. 2:3 covers the latter, but that is one option for one specific platform.
I genuinely do not understand the product decision. OpenAI has the compute. The previous-generation 4o image stack supported a wider range. Whatever the reason for the cap, this is a meaningful gap against Google, ByteDance, and even xAI on a single specific axis.
Where GPT Image 2 actually wins
Balancing the aspect ratio gripe: the model has real strengths over Nano Banana Pro in a few specific places.
Text rendering. Always the GPT Image lineup's strongest suit. GPT Image 2 keeps the family lead. If your image needs readable text (a poster, a sign, a UI mockup, a meme caption, a menu) GPT Image 2 is still the sharpest pick on the platform. Nano Banana Pro has gotten better at text but still produces the occasional "Esspresso $3.5O" garbage on longer copy. GPT Image 2 will give you legible eight-line menus.
Edit operations on a reference image. Feed it a portrait and tell it to swap the background while preserving the subject. Feed it a product shot and tell it to change the bottle color while keeping the typography intact. GPT Image 2 understands the boundary of "what to change" versus "what to preserve" cleanly. Nano Banana Pro does this well too, but GPT Image 2 is more obedient about not drifting elements you did not ask it to touch.
Instruction following on long prompts. Write a prompt with 8 to 10 distinct requirements and GPT Image 2 will respect more of them than 1.5 did. Not perfect, but the regression rate as prompt complexity climbs is genuinely better.
Structured compositions. Tables, charts, layouts with explicit positioning. GPT Image 2 handles spatial reasoning better than most alternatives. Mocking up a UI screen, an infographic skeleton, or a layout where the position of elements matters? This is your model.
How we are using it on Picturesque
A few concrete decisions we made this week:
We updated the template library so any template currently running on Nano Banana Pro is also available on GPT Image 2. Same prompts, same example imagery. Users pick which model they want to spend credits on. For roughly 70 percent of our templates the output is comparable enough that the choice comes down to budget.
We did not flip the default. Nano Banana Pro is still the recommended model for character work, brand consistency, and any template aimed at realistic portraiture. GPT Image 2 is the recommended model for templates aimed at illustration, posters with text, and structured layouts. For the overlap zone, it is a genuine user choice.
When a template is applied, the studio automatically selects high quality on GPT models so creators get the best version of their work without remembering to bump it. This was already true for resolution-tiered models like Nano Banana Pro. GPT Image 2 just slots into the same automation.
The honest verdict
GPT Image 2 is a solid model from a frontier AI company. The quality is now genuinely close to Nano Banana Pro on most prompts. Pricing is half. Text rendering and edit obedience are both meaningful advantages.
It is not a replacement for Nano Banana Pro in our workflow though. The polish problem (the AI Slop look on realistic work) and the three-aspect-ratio cap (a real limitation for social-first delivery) are both real, and both push us back toward Google's model when those things matter.
The right framing is not "which is better." It is "which is better for what." GPT Image 2 is the new default for editorial illustration, posters with text, structured layouts, and edit operations. Nano Banana Pro is the default for portraits, character sets, brand campaigns, and anything that needs to land on a non-standard aspect ratio. For everything in between, pick on price and pick on the aesthetic you actually want.
If you have been using GPT Image 1.5 as your daily driver, the upgrade to 2 is a clear win. Same price, noticeably better quality, same workflow.
If you are a Nano Banana Pro user, give GPT Image 2 a few generations on prompts where you do not strictly need 4K or unusual aspect ratios. You will probably keep Nano Banana Pro as your premium pick. You will probably also find GPT Image 2 sneaking into a chunk of your day-to-day work.
That is closer to the goalposts than I expected from a "version 2" release. If they fix the aspect ratio cap, the conversation gets a lot more interesting.