MiniMax-H3-Image vs GPT Image 2.5: Both Take 16 References, and That Is Where the Similarity Ends

The number sixteen appears on both spec sheets, and it means two different things. MiniMax-H3-Image accepts up to sixteen reference images in a single edit, addressed by upload order as <Picture 1> through <Picture 16>, with the request body capped by nothing except a 32,000-character prompt. GPT Image 2.5 — the Sunburst and Flare pair that OpenAI shipped on 8 September 2026 — accepts up to sixteen input images too, and its documentation is explicit about what happens when you send more than one: markup and masks apply to the first image, and everything after it is a reference. Sixteen slots against sixteen slots, one of which is the whole control surface and the other of which is a parameter you will probably never fill.

That coincidence is the most useful thing about putting these two models side by side, because everything else about them is asymmetric in a way that is easy to state and hard to overstate. GPT Image 2.5 is the second and fourth best image model in the world on Artificial Analysis's text-to-image board, with thirteen thousand blind votes behind each placement. MiniMax-H3-Image has no placement at all — its weights landed on Hugging Face on 10 November 2026 at 14:50 UTC, one day before this was written, and no arena has had time to score it. So this is not a comparison of two candidates on equal footing. It is a comparison of a measured model and an unmeasured one, and the useful question is what the unmeasured one offers that would justify finding out.

What the two calls actually look like

OpenAI's side is two model ids doing one job between them. gpt-image-2.5-sunburst is positioned for detailed compositions and precise editing; gpt-image-2.5-flare is the fast everyday variant. Both share the same snapshot date, the same endpoint surface and the same rate, and the split is latency against precision rather than capability against capability. There is no bare gpt-image-2.5 alias — you pick a variant. Generation goes to v1/images/generations, edits to v1/images/edits as multipart form data, and Batch is supported at half the output rate. Quality is a dial with six settings — low, medium, high, xhigh, max and auto, with auto the default.

MiniMax's side is one model with two service tiers. MiniMax-H3-Image is a 33.1B dense diffusion transformer — 50 layers, hidden size 5,376, 56 attention heads, FFN width 14,336, 3D multimodal RoPE — distilled to six sampling steps with no classifier-free guidance. The served API enforces that: num_inference_steps must be 6 and guidance_scale must be 1, or you get an HTTP 400, and negative prompts are refused. Highspeed is the same weights at a higher per-token rate. There is no quality dial, no step count to raise, no second variant with different behaviour. You send a prompt and up to sixteen pictures and you get one image back, in a fixed number of passes.

Read those two paragraphs next to each other and the design philosophies separate cleanly. OpenAI gives you knobs and expects you to find the setting that fits the job. MiniMax gives you one setting and expects the model to be good enough at it. The first is a tool; the second is an appliance.

Size is where the gap is widest, and it is not close

GPT Image 2.5 tops out at 3840×2160 — 8.29 megapixels — with dimensions required to be multiples of 16, an aspect ratio between 1:3 and 3:1, and anything above 2560×1440 flagged as experimental in OpenAI's own documentation. The recommended presets are 1024×1024, 1536×1024 and 1024×1536, and the custom-size path allows a pixel budget from 655,360 up to that 8,294,400 ceiling.

MiniMax-H3-Image's ceiling is roughly double that, and its floor is higher than OpenAI's ceiling-to-floor gap. The card claims about 16 megapixels at every aspect ratio — 4096×4096 at 1:1, 4736×3552 at 4:3, 5024×3360 at 3:2, 5472×3072 at 16:9, 6272×2688 at 21:9 — with the API capping output area at 20 megapixels and the practical working range described as 4 to 20 MP. There is no upscaler in the path; the model regenerates detail rather than enlarging it.

• Maximum output — GPT Image 2.5: 3840×2160, 8.29 MP. MiniMax-H3-Image: about 16 MP, up to 20 MP through the API.

• Minimum useful output — GPT Image 2.5: 655,360 pixels, so a 768-pixel thumbnail is a supported request. MiniMax-H3-Image: roughly 4 MP, with no small-image tier.

• Reference images — both up to 16; OpenAI types the first as the edit target and treats the rest as references, MiniMax orders all sixteen positionally with no typing.

• Quality control — GPT Image 2.5: six quality settings plus auto. MiniMax-H3-Image: none; six steps, no guidance, enforced.

• Output format — GPT Image 2.5: PNG by default, JPEG and WebP available, compression 0–100, transparent background via background: "transparent". MiniMax-H3-Image: a PNG in data[0].b64_json.

One of these differences decides more deployments than the others. If your deliverable is a 4K plate, GPT Image 2.5 cannot make it — not at a lower quality, not with a setting changed. It stops at 8.29 megapixels and the rest is upscaling. That is not a criticism of the model; it is a statement about what the product is for. But it does mean the two models stop competing the moment your requirement crosses 8.3 MP, and no amount of measured Elo closes that.

Both meter tokens, and only one publishes what a token buys

OpenAI bills GPT Image 2.5 by tokens: $5 per million text input tokens ($1.25 cached), $8 per million image input tokens ($2 cached), and $30 per million image output tokens, halved to $15 on the Batch API. Text output is not charged because the model returns images rather than prose. Those rates are unchanged from GPT Image 2, which is worth knowing if you priced the previous generation.

The number that matters and does not appear in OpenAI's rate card is how many output tokens an image costs, and OpenAI does not publish it — per-image consumption was not documented at launch and is still not documented per size. The only per-image figure in circulation is Artificial Analysis's derived $210.72 per thousand images, which is about $0.21 a frame at 1024×1024 with default settings. That is a third-party derivation from measured usage, not an OpenAI price, and it should be treated as an estimate that moves with quality setting and resolution.

MiniMax publishes the resolved number directly, because its output rate is low enough that the arithmetic is simple: $2.30 per million input text tokens, $1.50 per million input image tokens, $3.80 per million output image tokens. The vendor's own worked examples give $0.02179 for a 1K image, $0.02333 at 2K and $0.06879 at 4K — the last of which is $0.03439 while a 50% discount holds, and that discount expires on 17 November 2026 at 16:00 UTC and applies only when the delivered image is at least 16,600,000 pixels.

Put the two side by side and the shape of the trade is stark. At 1024×1024, roughly $0.21 against $0.02179 — about ten to one. At the sizes each model is designed for, the comparison gets harder to make honestly, because GPT Image 2.5 at 8.29 MP and MiniMax-H3-Image at 16 MP are not the same product. What can be said without hedging is that if you are generating at 1K in volume, MiniMax's price is an order of magnitude lower and its quality is unmeasured; that is the entire bet, stated plainly.

Two billing details on the MiniMax side are easy to miss. Input images are billed per token at $1.50 per million, so references are not free — going from one to three references lifts a 2K edit from $0.02658 to $0.04177 in the vendor's own example. And the API charges for a request that times out, so a client-side retry policy turns one generation into two or three billed generations. OpenAI's documented behaviour is the opposite for failed and moderated requests, which it does not charge for.

The evidence base is lopsided, and both halves of that matter

On Artificial Analysis's text-to-image board, GPT Image 2.5 Sunburst sits second at Elo 1228 over 13,370 samples and Flare fourth at 1209 over 12,695. On the editing board the same pair is third and fifth — 1199 over 18,617 samples for Sunburst, 1181 over 19,498 for Flare. Those numbers are point estimates with confidence intervals in the tens of Elo, which is to say Sunburst and Flare are not meaningfully separated from each other on text-to-image and both are meaningfully above everything below rank five.

It is also worth noting that GPT Image 2.5 is no longer the top of its own family. The 2.6 variants — Sunburst and Flare — now hold first and third on the text-to-image board and first and second on editing, at the same $210.72 per thousand images. So a reader choosing between MiniMax-H3-Image and GPT Image 2.5 is choosing between an unmeasured newcomer and the previous generation of a measured line, while the current generation of that line sits one model id away at the same price. Anyone who picks 2.5 for the score should look at what the score is actually relative to.

MiniMax-H3-Image has no score on either board and no vendor evaluation table — no GenEval, no DPG-Bench, no human preference numbers. The card's claim is qualitative: “the strongest instruction following among open-source models,” with no comparison table attached. The showcase grid is the only output evidence that exists, and a showcase grid is a curated artefact. None of this is suspicious for a model that had been public for a day when it was written. It does mean that every argument for it in this article is an argument from price, from resolution or from architecture, and none of them is an argument from measured quality.

Prompt budget, and what a model can hold in its head

GPT Image 2.5 accepts a normal OpenAI prompt and, per OpenAI's documentation, up to sixteen input images; the practical constraint on complex edits is not a character count but the mechanics of the request — the mask must match the first image in size and carry an alpha channel, and the annotation layer for markup edits must match the first image's dimensions exactly. There is no published token ceiling for image prompts in the model's documentation, which means the ceiling you plan against is the API's, not the model's.

MiniMax-H3-Image publishes a specific and modest budget: 32,000 characters on the served API, of which only about the first 4,096 tokens take effect, and the local SGLang path truncates at 4,096 tokens outright. For a model whose pitch is instruction following, four thousand tokens is tight — a detailed brief with sixteen reference images referenced by position eats a large fraction of it. MiniMax's answer is a companion model: MiniMax-H3-Image-PE, a 27B prompt expander that rewrites a short request into a long caption plus an aspect-ratio recommendation and emits it as JSON. It fits in about 19 GiB of VRAM. Its editing path is described as prompt-only, so it helps text-to-image today and not much else.

That asymmetry is worth naming because it changes what “prompt following” means for each. OpenAI's model takes the prompt you wrote. MiniMax's takes the prompt you wrote, or a prompt another model rewrote for it, and either way it has less room to hold the instruction. If your workflow depends on long, structured briefs with many named references, that is a real constraint and it will show up before quality ever does.

Licence, retention and the cost of being closed

GPT Image 2.5 is closed. There are no weights, no self-hosting, no checkpoint to pin, and the commercial terms are an API contract. What OpenAI's developer documentation does state is that API data is not used for training unless you opt in, that the images endpoints are listed as not used for training, and that the models are Zero Data Retention compatible. Ownership and commercial-use terms live in separate policies, not in the developer docs.

MiniMax-H3-Image is open, and its licence is more restrictive than the word implies. The MiniMax-H3-Image Community License Agreement excludes the European Union, the United Kingdom, the Republic of Korea and the United States by definition; requires separate written authorisation above one million US dollars of annualised revenue; requires the same for any hosted service with no revenue floor; requires the model's name displayed prominently on commercial interfaces; and forbids using outputs to train or improve other models. The Hugging Face page surfaces all of this as license: other.

So the usual framing inverts here. The closed model has the simpler terms and the weaker lock-in — you can leave, you just cannot take the weights. The open model has the harder terms and the stronger lock-in, because a licence that excludes four major markets is a constraint on your business that no amount of downloadable weights offsets. If a self-hosted path is the reason you are looking at MiniMax-H3-Image, read the licence before the model card.

When the unmeasured one is the right call

There are three cases where MiniMax-H3-Image is the obvious choice despite having no scoreboard, and none of them is “because it is cheaper.”

The first is resolution. If you need more than 8.29 megapixels out of a single generation, GPT Image 2.5 is out of the running and the comparison is over. Sixteen megapixels at $0.06879 against upscaling an 8.29 MP frame is not a close call.

The second is reproducibility. Six steps, no guidance, a published schedule in pdd_config.json, and weights you can download. If you need the same output from the same input on your own hardware a year from now, GPT Image 2.5 cannot offer that and MiniMax-H3-Image can, licence permitting.

The third is volume at small sizes. At $0.02179 against roughly $0.21 at 1K, a pipeline generating a hundred thousand thumbnails a month is looking at a difference measured in tens of thousands of dollars. That is enough money to justify running your own blind comparison for a week, which is the only way either of these numbers becomes a decision.

Everything else points the other way. If you need a model that will not surprise you, GPT Image 2.5 has two board placements with thirteen thousand votes each and six weeks of paying traffic behind them; MiniMax-H3-Image has a day. If you need a quality dial, OpenAI has six settings and MiniMax has none. If you need a small image, OpenAI goes down to 655,360 pixels and MiniMax does not go below about 4 MP. And if you need the current best in that family rather than the previous generation, the 2.6 variants are the same price.

OrcaRouter routes both GPT Image 2.5 variants — gpt-image-2.5-sunburst and gpt-image-2.5-flare — under one key, at provider list price with 0% markup passed through, so a cut or a retirement on OpenAI's side is reflected the day it happens rather than the next time you migrate a billing integration. We do not serve MiniMax-H3-Image; it has no route on our catalogue, and the honest path to it today is MiniMax's own API or the 135 GB of downloads. What we can do in a comparison like this is keep the measured side cheap to test and easy to fail over from, which is the part of the decision that does not depend on an arena score.

What would make this a real comparison

Two things, and both are weeks rather than months away. The first is a MiniMax-H3-Image entry on an image arena with a few thousand samples, which will either confirm the price is a bargain or show it as what it currently is — an untested discount. The second is a per-image token figure from OpenAI at 2K and 4K, because the $0.21 estimate is a 1K number and the whole price argument shifts if the 4K multiplier is not what people assume.

Until both exist, the decision is unusually clean: buy the measured model for work that has to ship at moderate resolution, and treat MiniMax-H3-Image as a candidate for the jobs GPT Image 2.5 structurally cannot do — sixteen megapixels, exact reproducibility, or volume where the ten-to-one price gap pays for the evaluation several times over.