Updated 12 November 2026. Prices and licence terms in this space change fast. I re-check this page monthly.
Short answer: MiniMax H3-Image is the cheapest serious 4K image model you can buy by the image this week. Through the API a 4K frame costs $0.06879, against $0.113 for Nano Banana 2.1 and $0.24 for Nano Banana Pro. Use it for drafts and test sets through the API. Do not download the open weights for brand work until someone has read the licence. Above $1 million in yearly revenue it needs MiniMax’s written permission. It does not apply at all in the US, the EU, the UK or South Korea. And it asks you to label every public image as machine-made.
MiniMax H3-Image and Nano Banana in one table
| MiniMax H3-Image | Nano Banana 2.1 | Nano Banana Pro | |
|---|---|---|---|
| Released | 11 Nov 2026 | 6 Oct 2026 | earlier, Gemini 3 Pro Image |
| API price per image, 1K / 2K / 4K (standard) | $0.02179 / $0.02333 / $0.06879 | $0.0336 / $0.0504 / $0.113 | $0.134 / $0.134 / $0.24 |
| Native output | about 16 MP at every aspect ratio, 4096×4096 to 6272×2688, no upscaler | up to 4K | up to 4K |
| Reference images | up to 16 | yes (set consistency is its strength) | yes |
| Free route | a demo Space on Hugging Face | no (Nano Banana 2 is the free one in the Gemini app) | no |
| Weights you can download | yes, under a community licence | no | no |
MiniMax prices are from its API pricing page; until 16:00 UTC on 17 November the 4K tier is half price ($0.03439). Real cost moves with prompt length, because MiniMax also bills the tokens of its own prompt expansion. Google prices are from the Gemini API pricing page. Both read 12 November 2026. Model details are from the H3-Image model card.
What MiniMax claims, and what it does not show
MiniMax says a 4K image takes under ten seconds and that its high-speed mode is three to four times faster than "mainstream alternatives", which it leaves unnamed (MiniMax, 11 Nov). There is no independent benchmark yet. One measured number exists: the vLLM team timed a 4736×3552 image at 3.84 seconds on four H200 GPUs (vLLM recipe).
Open weights here does not mean your laptop. The image transformer has 33.1 billion parameters. The model and its text encoder are about 135 GB together, and running everything on one GPU needs an estimated 135 GiB of memory. That is data-centre hardware (model card, recipe). The model expects a detailed English prompt. The API rewrites a short prompt in any language into one. The showcase proves on-image text in Chinese and English only.
What the licence actually says
I read the community licence (10 November 2026) before anything else, because an in-house team runs into four of its lines on day one.
If you and your affiliates earn more than $1 million a year, you need MiniMax’s prior written authorization, and every FMCG group I know is past that line. The licence covers the world except the European Union, the United Kingdom, the United States and South Korea, and the outputs may not be used in those places either. Its use policy forbids publishing machine-generated images in any public setting without clearly saying so, which covers every ad. And you may not use its outputs to improve any other AI model.
MiniMax claims no rights over what you generate. The hosted API runs under MiniMax’s separate platform terms, which this page does not cover. Read them before a paid job.
My read: cheap is a test budget, not a production decision
I work in-house at a large FMCG group, so the revenue line is not theoretical for me. I have not run H3-Image on client work yet, and this page is about what you can know before the first test.
Price matters at the draft stage. On client films I reject 60 to 70 percent of my takes before anyone else sees them. When most frames go in the bin, a cheaper frame is a real saving. The cover image on every AI Daily on this site comes from the Gemini app. A board made there carries Google’s SynthID watermark and a visible one (Google). It is free, and nobody mistakes it for a final.
The final is where the licence bites. On the Festival Toffee 3D ad, every shot was checked against a product sheet I built first. With H3-Image I would add one more document to that folder: the written authorization, and the disclosure line the media plan will carry. If those two are missing, the cheaper frame is not cheaper.
Open weights are not a free licence. MiniMax H3-Image costs about 60 percent of Nano Banana 2.1 at 4K, but a company earning over $1 million a year needs MiniMax’s written permission, and every public image must be labelled as machine-made.
Questions people are asking this week
Is MiniMax H3-Image better than Nano Banana?
Nobody has shown that yet. MiniMax claims speed and prompt adherence, but names no rival and publishes no independent test. Nano Banana 2.1 still has Google’s published scores on product consistency across a set. Compare them on your own brief. My Nano Banana comparison covers the Google side.
Is MiniMax H3-Image free?
The weights are free to download, and MiniMax runs a demo Space on Hugging Face. Commercial use above $1 million in revenue needs written permission, and the API is paid per image (MiniMax pricing).
Can I use it for ads in the US or Europe?
Not under the open-weights licence. The US, the EU, the UK and South Korea are excluded, for the model and its outputs (licence §V.4 and Exhibit A). MiniMax says it will license those regions case by case.
How do I write a good prompt for it?
Decide the shot before you type: subject, framing, light, palette and what must not appear. Write it in English if you run the weights yourself. My free AI image prompt generator walks through those decisions.
What I’d do this week
Before 17 November, while 4K is half price, take one brief you have already shipped. Run a six-frame set through the H3-Image API at 2K next to Nano Banana 2.1, from the same real pack shot. Count the frames you would reject in each. Then send the licence to legal before any frame goes public. If you would rather have the set directed for you, this is what I do.
Sources
- MiniMax — MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era, 11 Nov 2026 — https://www.minimax.io/blog/minimax-h3-image
- MiniMax — MiniMax-H3-Image model card (sizes, 33.1B DiT, prompt language, download sizes), read 12 Nov 2026 — https://huggingface.co/MiniMaxAI/MiniMax-H3-Image
- MiniMax — MiniMax-H3-Image Community License Agreement and Acceptable Use Policy, 10 Nov 2026 — https://huggingface.co/MiniMaxAI/MiniMax-H3-Image/blob/main/LICENSE
- MiniMax — API pricing, image generation (standard, highspeed, 4K launch discount to 17 Nov), read 12 Nov 2026 — https://platform.minimax.io/docs/guides/pricing
- vLLM — MiniMax-H3-Image deployment recipe (memory, measured latency), read 12 Nov 2026 — https://recipes.vllm.ai/MiniMaxAI/MiniMax-H3-Image
- Google — Gemini API pricing, image models, read 12 Nov 2026 — https://aistudio.google.com/docs/pricing
- Google — Gemini app image generation (free vs paid models, SynthID and visible watermark), read 12 Nov 2026 — https://gemini.google/overview/image-generation/
