MiniMax H3 Image is MiniMax's image model, released November 10, 2026. In our tests it was close to GPT Image 2.6 in quality and different in character: it adds atmosphere and details of its own, where GPT Image 2.6 does exactly what the prompt says. A picture costs 6 credits, less than GPT Image 2.6 at High (8 to 12) and more than at Low on a wide frame (4).

Tested November 12, 2026: 61 pictures on fal, the same prompts on H3 Image, GPT Image 2.6 Flare and Nano Banana Pro, at the size Tonpit sells each model. They cost $2.88 at fal's rates. Prices below are in credits as Tonpit charges them; a credit is one cent.

An elderly fisherman in a navy sweater by a wooden window, looking out at a harbour and a lighthouse, his hand on a coil of rope — MiniMax H3 Image: 6 credits.

An elderly fisherman in a navy sweater facing the camera, both hands on a coil of rope, fishing nets behind him — GPT Image 2.6 Flare, High: 12 credits.

An elderly fisherman in a navy sweater seated in an armchair by a window, hands on a coil of rope — Nano Banana Pro: 30 credits.

01 · In short

What is MiniMax H3 Image?

A model for pictures and edits. It makes a picture from text, or edits one from up to 16 input pictures.
2K and 4K in Tonpit. It draws every picture at no less than the area of a 2K square, so a smaller one would cost the same as 2K. We offer no 1080p.
It rewrites your prompt into a long description before it draws. You can switch that off for a picture from text; an edit always rewrites.
No seed, no mask, no transparent background. GPT Image 2.6 has masks, markup edits and transparency.

02 · Compared

How does it compare to GPT Image 2.6 and Nano Banana Pro?

Close in quality, different in character. H3 Image adds story to a prompt (a lighthouse in the window, beans around a bottle) and goes for dramatic framing; GPT Image 2.6 is literal and exact; Nano Banana Pro tends to show more of the scene. Each row is one prompt run on every model.

Test	H3 Image	GPT Image 2.6	Nano Banana Pro
Fisherman portrait	Level	Level, at Low and High	Level
Bottle with a 3-line label	Spelled right, added coffee beans	Spelled right	Stray apostrophe
Chalkboard menu, 8 lines	Spelled right	Spelled right, best layout	Spelled right
Edit: swap a lamp, add an apple	Right	High right, Low half right	Right
Night market crowd	Level	Level	Level
Two people from references	Both kept	Both kept	Both kept
Five objects placed exactly	One misplaced	All right	Not run
Alpine village at 4K	Missed the golden light, rewrite off	Followed the brief	Not run

03 · Price and speed

What does MiniMax H3 Image cost, and how fast is it?

6 credits a picture up to 2K, 4 with the rewrite off, 9 at 4K in a wide frame, and each picture you give it adds 1 to 3 credits by its size. A picture from text took 9 to 15 seconds, an edit 15 to 26.

Model	Square	Wide, 16:9	4K, 16:9
MiniMax H3 Image	6	6	9
GPT Image 2.6 Flare, Low	6	4	10
GPT Image 2.6 Flare, High	12	8	21
Nano Banana Pro	30	30	60

Speed is close. Once OpenAI's service had recovered, GPT Image 2.6 at High took 10 to 23 seconds a picture, against H3's 9 to 15. Nano Banana Pro took 18 to 29.

Every size up to 2K costs the same because the model draws at the same area whatever you order: a 1088×1088 picture billed 4,096 tokens and a 2560×1440 one 4,080. 4K bills about twice as many.

A second supplier matters more than it sounds. On the afternoon we tested, GPT Image 2.6 failed 33 of 51 requests with errors from OpenAI's side, some of the rest took close to two minutes, and H3 Image and Nano Banana Pro failed none. One afternoon is not a pattern, but it is the day a fallback earns its place.

04 · Edits

How good are its edits?

Good: it changes what you ask and keeps the rest, three edits in a row included. We asked each model to replace the brass desk lamp with a green vase of white tulips, add a red apple by the mug, and change nothing else.

A sunlit home office with a desk, a monitor, a brass desk lamp and a mug, an open window over white buildings and the sea — The picture we edited.

The same office with a green ceramic vase of white tulips where the lamp stood and a red apple by the mug — H3 Image: both changes, the room kept. 7 credits.

The same office with a clear glass vase of tulips in front of the lamp, which is still there, and an apple by the mug — GPT Image 2.6 Low: a glass vase, and the lamp left standing.

GPT Image 2.6 at High and Nano Banana Pro made the same edit correctly. Then we edited the same room on H3 Image three times in a row, each edit on the last one's result: night, then snow outside, then a cat on the chair. The room held through all three.

The office at night, lit only by the desk lamp, the city outside dark with lit windows — Edit 1: make it night.

The same night office, with snow on the rooftops and palms outside the window — Edit 2: snow outside.

The same snowy night office with a grey cat asleep on the woven chair — Edit 3: a cat on the chair.

05 · Consistency

Can it keep the same person across pictures?

Yes. We gave it one portrait of Rúna, one of Tonpit's ready-made fictional characters, and asked for three new scenes. Her face, red braids and patterned sweater held in all three, which is what you want from frames that will become a video.

A red-haired woman with braids in a patterned sweater steering a wooden boat through a grey, choppy sea — On a boat.

The same woman frying fish in a small kitchen, steam rising from the pan — In a kitchen.

The same woman selling fish on ice at a harbour market, talking to a customer — At a market.

06 · The rewrite

What does the prompt rewrite do, and when should you turn it off?

It turns your prompt into a long description before the model draws: that places things better, and it adds things you did not ask for. Asked for three red cubes with a yellow cone on the middle one, the model put the cone on the end cube with the rewrite off and on the middle one with it on. Asked for a bottle on slate, the rewrite added a cork and coffee beans.

An amber bottle with a cork and a cream label reading NORDLYS, Cold Brew Coffee, 330 ml, coffee beans scattered on wet slate — Rewrite on: a cork and coffee beans nobody asked for. 6 credits.

An amber bottle with a cap and the same label on dark slate, nothing else in the frame — Rewrite off: the prompt as written. 4 credits.

Turn it off for a label, a menu or a product that must carry nothing extra, and leave it on for a scene you have not fully described. The text you quote survives the rewrite word for word.

07 · Precision

What does GPT Image 2.6 do differently?

It is more literal, which shows when a picture must match the brief to the object. In the cubes test, GPT Image 2.6 put every object where it was asked at Low and at High; H3 Image laid the green ring between the spheres instead of against the right one. At 4K, GPT followed the brief for an alpine village better (golden light, the cows, the bridge), though we ran H3 with its rewrite off there. GPT cost 21 credits, H3 7.

Three red cubes with a yellow cone on the middle one, two blue spheres, and a green ring lying between the spheres — H3 Image: the ring between the spheres.

Three red cubes with a yellow cone on the middle one, two blue spheres, and a green ring leaning against the right sphere — GPT Image 2.6 High: every object in place.

08 · Choosing

Which image model should you use?

There is no best image model for everything. Each has its own character, so when a picture is good but not the one you had in mind, run the same prompt on another model before rewriting it. Reach for H3 Image for mood and story, and when you want a second supplier beside OpenAI; for GPT Image 2.6 when every object has to land exactly.