Both take sixteen references — the highest on the image roster — and both use a quality toggle rather than resolution tiers. From there they diverge sharply. GPT Image 2 offers Auto, Low, Medium and High plus nine aspect ratios plus a 1–10 variations field. GPT Image 1 Mini offers Auto and Low, four aspect ratios, and **no variations field at all** — despite a description promising ten at a time.
The Mini tier is described as the batching model. Its panel has no batching control. That's worth knowing before you plan a workflow around it.
What differs?
Three rows out of five, and one of them contradicts a description.
| GPT Image 2 | GPT Image 1 Mini | |
|---|---|---|
| Quality | Auto / Low / **Medium / High** | Auto / Low only |
| Aspect ratios | **9**, including auto | **4**, including auto |
| Variations | **1–10** | **None** |
| References | 16 | 16 |
| Accepted types | Images | Images |
Source: Hexcoded Creative Studio picker, Image → Create, September 2026.
Take the identical rows first. Both take sixteen references, which is the highest ceiling on the image roster — nothing else exceeds fourteen. And both accept image references only.
Then the three that differ.
Quality. GPT Image 2 has the full four-step toggle. Mini has Auto and Low, with no Medium or High. So Mini can't reach the top two quality settings at all.
Aspect ratios. Nine against four. Mini offers 1:1, 2:3, 3:2 and auto — the narrowest range on the entire image roster.
Variations. GPT Image 2 has a 1–10 field. Mini has none.
GPT Image 1 Mini's description says "rapid concept batches, ten at a time." Its panel has no variations field. GPT Image 2's does, at 1–10.
Where this falls short. That's what the picker shows. Whether Mini batches by some other mechanism the panel doesn't expose, or whether the description is simply describing the wrong model, isn't something we can determine from the interface. What's verifiable is that the control isn't there.
Why do these two use quality instead of resolution?
Because they're the only OpenAI models on the image roster, and they follow a different convention.
Every other image model uses resolution tiers in K — 1K, 2K, 4K. These two use a quality toggle: Auto, Low, Medium, High.
That's not a small difference when comparing across the roster. A quality setting and a resolution tier aren't the same kind of thing, and there's no stated mapping between them. "High" doesn't correspond to a stated pixel dimension anywhere the picker exposes.
Which means comparing GPT Image 2's High against Nano Banana Pro's 4K isn't a comparison you can make from the settings.
What's the sixteen-reference ceiling for?
It's the highest on the image roster, and it's worth noting how far ahead it is.
Sixteen on both OpenAI models. Fourteen on the Nano Banana family and Seedream 4.5. Ten on Seedream 5.0 Pro. Eight on FLUX.2 Max and Pro. Four on FLUX.2 Klein 4B. One on Recraft 4.1.
Hexcoded's own documentation gives sixteen as the image reference pool, which reads as a platform ceiling. In practice only these two models offer it — so sixteen is the maximum rather than the standard.
For reference-heavy work, that makes these two the obvious starting point regardless of anything else in the comparison.
What is GPT Image 2 for?
Text and prompt precision, by its own description.
"Exacting prompt control and clean in-image text." It's one of only two image models whose description mentions text rendering — the other is Seedream 5.0 Pro, which names multilingual text.
So for anything where copy appears inside the image — a poster, a title card, a product graphic with a label — these two are the models whose descriptions claim it.
Where this falls short. A description naming a capability isn't a measurement of it. We haven't tested in-image text rendering on either model, and the picker exposes nothing about it. For anything where the text is load-bearing, test it on your own copy.
What do they cost?
Only GPT Image 2 is on the published panel, and it's the most expensive image model listed.
70 credits for 1K, 100 for 2K, 137 for 4K. For comparison the panel lists Seedream at 15 for 2K, Nano Banana Pro at 45 for 2K, and Recraft 4.1 at 12 standard.
GPT Image 1 Mini isn't listed at all.
Two things make those figures awkward to use. The panel prices GPT Image 2 in K tiers, and the picker doesn't offer K tiers on this model — it offers Auto, Low, Medium, High. So the mapping between the two is unstated.
Where this falls short. Given that mismatch, treat the figures as indicative of where GPT Image 2 sits relative to other models rather than as a price per quality setting. The cost shown before you render is the authoritative one.
The cost panel prices GPT Image 2 in K tiers. The picker offers Auto to High. There's no stated mapping between them.
So which should you use?
Three cases.
Anything with text inside the image
GPT Image 2, whose description names clean in-image text. Test it on your own copy first, since the claim isn't measurable from the settings.
Anything needing more than four aspect ratios
GPT Image 2. Mini offers 1:1, 2:3, 3:2 and auto — no 16:9, no 9:16, no 4:3. For vertical work that rules it out entirely.
Reference-heavy work at low quality
Either. Both take sixteen references, the highest on the roster. Mini's Auto and Low tiers are presumably cheaper, though it isn't on the published panel.
All settings read from Hexcoded's Creative Studio picker in Image → Create mode in September 2026, both models observed individually. GPT Image 1 Mini does not appear on the published render costs panel, and the panel prices GPT Image 2 in resolution tiers the picker doesn't offer for it. Settings and costs change with model versions. The cost shown before you render is the authoritative figure.
- Both take sixteen references — the highest on the image roster, and the only two that reach it
- Both use a quality toggle rather than resolution tiers. No other image model does
- GPT Image 2 offers Auto, Low, Medium and High. Mini offers Auto and Low only
- GPT Image 2 has a 1–10 variations field. Mini has none, despite a description promising ten at a time
- Mini offers four aspect ratios — 1:1, 2:3, 3:2 and auto. No 9:16, which rules it out for vertical work
- GPT Image 2's description names clean in-image text. One of only two image models that mention text
- GPT Image 2 is the most expensive image model on the published panel. Mini isn't listed
- The panel prices GPT Image 2 in K tiers the picker doesn't offer. No stated mapping between them
Three things. GPT Image 2 has the full Auto to High quality toggle, nine aspect ratios and a 1–10 variations field. Mini has Auto and Low only, four aspect ratios, and no variations field. Both take sixteen references.
Its description says "rapid concept batches, ten at a time," but its panel has no variations field. GPT Image 2 has one, at 1–10. So the batching control sits on the other model, whatever the description says.
They're the only OpenAI models on the image roster and they follow a different convention. Every other model uses resolution tiers in K. There's no stated mapping between the two systems, so "High" doesn't correspond to a published pixel dimension.
Not at 9:16. Its four aspect ratios are 1:1, 2:3, 3:2 and auto — so 2:3 is the closest to portrait available. It's the only image model on the roster without a 9:16 option, which rules it out for most short-form work.
GPT Image 2's description names it — "exacting prompt control and clean in-image text." It's one of only two image models whose description mentions text, alongside Seedream 5.0 Pro's multilingual claim. Neither claim is measurable from the settings, so test it on your own copy.
Sixteen each, which is the highest on the image roster. The next tier is fourteen on the Nano Banana family and Seedream 4.5. Hexcoded's documentation gives sixteen as the image reference pool, but only these two models actually offer it.
Sixteen references, two quality systems
Reference ceilings, quality tiers, aspect ratios and batching all visible per model before you generate. Eleven image models on the same credit balance as the video roster.
Open Creative StudioMore on model capability, access and rights in Models.