MiniMax-H3-Image vs GPT Image 2.6: What You Can Actually Change in Each Request

Most image-model comparisons spend their space on scores. This one spends it on knobs, because MiniMax-H3-Image and G​PT Image 2.6 Sunburst are not competing over the same dial. MiniMax-H3-Image is an open-weight 33.1B diffusion transformer, released on 10 November 2026, that runs a fixed six-step sampler with no guidance scale, accepts up to sixteen reference images and generates up to sixteen megapixels natively. G​PT Image 2.6 Sunburst — and its faster sibling G​PT Image 2.6 Flare, both shipped 4 November 2026 — is a closed hosted model whose recent gains are mostly about what you can hand it and what it hands back: reused reference images, markup-guided edits, GIF and SVG output. If your bottleneck is controlling a long edit chain, one of these is obviously right. If your bottleneck is a resolution or a bill, the other is.

Both are worth taking seriously. They just fail in different directions, and the failure modes are legible from the control surfaces alone.

The sampler is a dial on one side and a constant on the other

MiniMax-H3-Image gives you almost nothing to tune, and enforces that at the API boundary. The checkpoint is CFG-distilled down to six function evaluations, which means the usual two knobs of diffusion work — step count and guidance scale — have been collapsed into the weights. Call it with num_inference_steps set to anything other than six and the request comes back as HTTP 400. Set guidance_scale to anything other than one and it is rejected the same way. Send a negative_prompt and it is refused outright, because there is no second forward pass for it to act on.

That sounds restrictive and is mostly a simplification. Six steps is fast, deterministic given a seed, and impossible to misconfigure — there is no "I turned the steps up and it got worse" failure. What you lose is the ability to trade quality for latency inside the model, and any negative-prompt trick you have accumulated.

G​PT Image 2.6 keeps its own small control surface in place of sampler knobs: a complexity parameter that defaults to high, which is the closest thing either model has to a "think harder" switch. Its documented additions for 2.6 are about inputs and outputs, not sampling — reference images that persist across follow-up edits so you stop re-uploading them, markup drawn onto a reference image to direct an edit, and two output formats the previous generation did not offer.

• Step count — MiniMax-H3-Image fixed at six, rejected otherwise; G​PT Image 2.6 has no exposed step control.

• Guidance — MiniMax-H3-Image pins guidance_scale to one; G​PT Image 2.6 exposes complexity (default high) instead.

• Negative prompts — refused with a 400 by MiniMax-H3-Image; not a documented surface on G​PT Image 2.6.

• Prompt budget — MiniMax-H3-Image truncates at 4,096 tokens; G​PT Image 2.6's prompt ceiling is not published on its model pages.

Sixteen reference slots is the widest control either model offers

The single biggest functional difference is how much context you can put in front of the edit. MiniMax-H3-Image accepts up to sixteen reference images in one request, and the model card frames the editing mode as working at roughly the same fidelity as generation — the vLLM-Omni port reports 0.985 SSIM and about 35 dB PSNR against its own single-GPU reference on a 2048² edit, which is a self-consistency check rather than a quality score, but it does mean the parallelised path is not quietly degrading your output.

Sixteen references is a different order of control from "one image plus a sentence". Product photography with six angles, a style sheet, a character sheet and a brand guideline can all go in at once. The cost of that breadth is that prompt engineering becomes context engineering: with sixteen images in play, the text has to say which one governs which attribute, and the model has no way to ask.

G​PT Image 2.6's answer to the same problem is memory rather than width. Reference images persist across follow-up edits inside a session, so a multi-turn retouch does not re-upload the original each time, and the markup-guided edit path lets you point at a region instead of describing it. For iterative work on one image, that is a better interface than sixteen static slots. For a one-shot composite built from a dozen assets, it is not.

What comes back out

The output format list is where G​PT Image 2.6 pulls ahead of every open-weight image model, including this one. Version 2.6 added animated GIF output and SVG output alongside the usual raster formats. An SVG result matters more than it sounds: it is the difference between an image model you use to draft a logo concept and one you can hand to a front-end. MiniMax-H3-Image returns PNG — its serving path is validated on bit-identical PNG comparison — and nothing in the vLLM-Omni pipeline or the model card suggests vector or animation output.

Resolution runs the other way, and it is not close. MiniMax-H3-Image generates up to sixteen megapixels natively, at every aspect ratio from a square to an ultra-wide, with no upscaler in the path — 4096×4096 and 6272×2688 are both native outputs. The working range is about four to twenty megapixels, with quality peaking near the top of it. G​PT Image 2.6's model pages document formats and reference handling but do not publish a maximum pixel count, so the resolution argument in this matchup is M​iniMax publishing a number against O​penAI not publishing one, rather than a measured head-to-head.

• Native output — MiniMax-H3-Image up to 16 MP at any ratio, no upscaler; G​PT Image 2.6 documents no maximum.

• Formats — MiniMax-H3-Image returns PNG; G​PT Image 2.6 adds GIF and SVG.

• Aspect ratios — MiniMax-H3-Image accepts any width and height pair; G​PT Image 2.6's set is not published per model page.

• Editing input — MiniMax-H3-Image takes up to 16 reference images per request; G​PT Image 2.6 keeps references across turns and accepts drawn markup.

The bill, on each vendor's own worked examples

Both vendors meter images as tokens, which makes the sticker prices look comparable and the per-image costs look nothing alike. M​iniMax publishes worked examples on its pricing page, so this is arithmetic on their own numbers rather than a derivation:

• MiniMax-H3-Image 1024×1024, text-to-image — $0.02179 per image, against $3.80 per million output tokens.

• MiniMax-H3-Image 2048×2048 — $0.02333 per image. Doubling the pixels costs half a cent, because token count grows far more slowly than area at this end of the range.

• MiniMax-H3-Image 4096×4096 — $0.06879 per image, cut to $0.03439 under a limited-time 50% offer that runs to 17 November 2026 at 16:00 UTC and requires the delivered image to be at least 16,600,000 pixels.

• MiniMax-H3-Image edit, one reference at 1024×1024 — $0.02654 per image, with input images billed at $1.50 per million tokens and input text at $2.30.

• G​PT Image 2.6 — output tokens at $30 per million. At the token count O​penAI's own documentation implies for a 1024×1024 result, that lands near 21 cents per image, roughly ten times the M​iniMax figure for the same nominal size.

The speeds differ too, and this is where the two products stop being interchangeable. On the numbers M​iniMax published with the 4 November release, G​PT Image 2.6 Flare runs a text-to-image request at a median around 30 seconds, down from roughly 55 for the prior generation, and an edit at about 32 seconds, down from roughly 58. Sunburst is the slower, higher-fidelity tier and got slower still — around 135 seconds against 113. MiniMax-H3-Image's own latency is only measurable where you host it: 0.88 seconds for a 2048² image on four H200s, 1.35 seconds for a 2048² edit, per the official recipe. Those are not the same kind of number — self-hosted node latency versus a managed endpoint's queue-inclusive median — and putting them in one sentence as if they were would be dishonest. What the comparison legitimately says is that a hosted 2.6 request takes tens of seconds and a local H3-Image request takes under two, provided you already own the four GPUs.

What the public evidence covers, and what it does not

Artificial Analysis scores both of them, on the same two boards, which makes this one of the few places in the matchup where the two sides are directly comparable rather than argued from spec sheets. On AA-Image-Editing v2.1, G​PT Image 2.6 Sunburst sits at 1,210 Elo over roughly 6,400 samples, Flare at 1,206 over some 6,700, and MiniMax-H3-Image at 1,138 in a rank band of eight to seventeen over about 3,200 samples. On AA-Image-T2I v2.1 the gap widens: Sunburst at 1,261 in first place over 5,282 samples against MiniMax-H3-Image at 1,066, ranked twenty-third. Same vote pool, same Elo scale, roughly seventy points apart on editing and nearly two hundred on generation.

Read that honestly and it says two things at once. MiniMax-H3-Image is the highest-ranked open-weight model on the editing board — ahead of Fibo 2.0 at 1,085 and Qwen-Image-2.1 at 1,084 — so its quality claim is no longer an assertion; it is a measured number. But the number is about seventy Elo below G​PT Image 2.6 Sunburst, and its interval spans ten ranks while Sunburst's top-three band does not. On the axis the two models most obviously share, G​PT Image 2.6 wins, and it wins by more on the board where H3-Image is weaker. The 16-megapixel resolution argument and the two-cent price still stand on their own; they just are not quality arguments.

What neither board tells you is how much of the G​PT Image 2.6 edge survives when the task is the one this article is about — a long edit chain with many references and a fixed sampler. Elo is an aggregate over human preference on unknown prompts, and a model can be seventy points ahead on that aggregate while losing on a specific workflow. The only way to find out is to run your own prompts through both, which is the recommendation at the end of this piece rather than a hedge in the middle of it.

Running them side by side without a second contract

The practical complication is that neither model is one API call away for most teams. MiniMax-H3-Image's day-one path is self-hosted — a 67B total assembly of a 33.1B transformer, a 33.4B Qwen3-VL text encoder reused unchanged from the video model, and a 0.58B VAE, with the 16-megapixel sweet spot needing four datacentre GPUs at full speed. It is also available through the vendor's own API and several third-party platforms under the token pricing above. G​PT Image 2.6 is O​penAI-only.

If you want to A/B the two on your own prompts, the honest description of the routing layer is this: OrcaRouter does not carry MiniMax-H3-Image yet, and does not carry G​PT Image 2.6 Sunburst or Flare. What it does carry is the previous G​PT Image pair — gpt-image-2.5-sunburst and gpt-image-2.5-flare — alongside gpt-image-2 and gpt-image-1.5, all at the provider's list price with no markup passed through on top. That makes a useful intermediate test: build the harness against 2.5 through one key, confirm your evaluation actually discriminates between Sunburst's fidelity and Flare's speed, and only then decide whether the 2.6 delta or a self-hosted open-weight model is the thing you need. Model fusion is the other option worth knowing about here — a panel of models answering the same image prompt gives you a spread of outputs per prompt without you writing three clients.

The reverse case is already covered, incidentally: minimax/minimax-h3, the video model whose DiT and text encoder MiniMax-H3-Image is built from, is routable today. If your reason for looking at the image model is that you already run H3 video, the family already answers on one key.

Which one to pick

Choose MiniMax-H3-Image when the constraint is pixels, references or unit cost: a 16-megapixel native output, sixteen input images, and roughly two cents for a 1024² generation. Choose G​PT Image 2.6 when the constraint is the shape of the output or the smoothness of an edit loop: SVG and GIF on the way out, persistent references and drawn markup on the way in, and a seventy-Elo lead on the editing board with a top-three interval behind it.

Choose neither with confidence if what you actually needed was a score for your prompts. Both models now have public scores, and both scores are aggregates over somebody else's prompts; the comparison that decides your case is the one you run yourself, with a bill attached to each side of it. What the leaderboard buys you is a prior — G​PT Image 2.6 is ahead, and on text-to-image it is ahead by enough that you should expect to see it. What it cannot buy you is the answer for a sixteen-reference product composite at four megapixels, because nobody has published that comparison and the Elo boards do not measure it.