MiniMax-H3-Image vs Nano Banana 2.1: Nearly the Same Price, Nowhere Near the Same Evidence

Two image models, both around three cents a picture, and the decision between MiniMax-H3-Image and Nano Banana 2.1 is not settled by the price. At 1024×1024 the open-weight newcomer costs $0.02179 a generation and G​oogle's Nano Banana 2.1 costs $0.0336 — a difference of about a cent. At 4096×4096 the gap widens to $0.06879 against $0.113. So the cheap one is genuinely cheaper, by a third to a half. What you cannot buy at any price right now is a reason to believe the cheap one's quality claim, because nobody outside M​iniMax has measured it. Nano Banana 2.1 sits fourth on both of Artificial Analysis's image boards with thousands of samples behind each rank. MiniMax-H3-Image appears on neither board, in any position, with any sample count. That asymmetry is the whole comparison, and no amount of spec-sheet arithmetic closes it.

Worth being concrete about dates before going further, because the two models are not in the same phase of life. Nano Banana 2.1 has been generally available since 6 October 2026 as gemini-nano-banana-2.1, and it replaced gemini-3.1-flash-image, which G​oogle deprecated with a shutdown date of 29 October 2026 — so if you were calling the older one, you have already had to move. MiniMax-H3-Image's weights landed on Hugging Face on 10 November 2026 at 14:50 UTC, with a terse card in M​iniMax's API release notes dated the same day and no accompanying blog post. One model has a month of production traffic; the other has a day of being downloadable. Judge accordingly.

What each one actually hands you

Nano Banana 2.1 is a Gemini 3.6 Flash derivative sold only through G​oogle's API. Three resolution tiers — 1K, 2K, 4K, and no 512-pixel option — with aspect ratios from the usual set, a 131,072-token input context and 32,768 tokens of output. It accepts up to 14 reference images in a single request, split by G​oogle into 10 object references and 4 character references, and it can ground a generation in G​oogle Search or Image Search results. Thinking is on by default at medium, with minimal and high available, and it cannot be switched off through the API at all. Every output carries a SynthID watermark.

MiniMax-H3-Image is a 33.1B dense diffusion transformer with a CNN encoder and a 24-layer ViT decoder for its VAE, shipped under M​iniMax's own community licence. Text-to-image and edit both run through one endpoint pair, edits take 1 to 16 reference images addressed positionally as <Picture 1> through <Picture 16> in upload order, and the model is distilled to six sampling steps with no classifier-free guidance — the served API returns HTTP 400 if you pass anything other than num_inference_steps=6 and guidance_scale=1. Output area is capped at 20,000,000 pixels through the API, with area tiers at 1K, 2K and 4K and 4K as the default; the card claims roughly 16 megapixels at every aspect ratio. There is no search grounding, no thinking budget and no watermarking scheme stated anywhere in the repository or the licence.

The capability lists diverge in a way that maps cleanly onto who each product is for. G​oogle shipped orchestration: references, retrieval, a reasoning budget, provenance marks. M​iniMax shipped resolution and control: sixteen reference slots, a 20-megapixel ceiling, a fixed six-step sampler you can replicate exactly on your own hardware.

The price ladder, with the discount window spelled out

Nano Banana 2.1 is metered at $30 per million output image tokens, which G​oogle's own documentation resolves to $0.0336 per 1K image, $0.0504 per 2K and $0.113 per 4K. Input images bill at $1.50 per million tokens and text output at $7.50 per million. There is a batch tier at half those output rates for jobs that can wait, and Artificial Analysis lists the effective 1K rate as $33.60 per thousand images — the same number expressed in bulk.

MiniMax-H3-Image is metered the same way in principle and very differently in outcome. Standard tier: $2.30 per million input text tokens, $1.50 per million input image tokens, $3.80 per million output image tokens. Worked through the vendor's own examples:

• 1K text-to-image — Nano Banana 2.1 $0.0336 vs MiniMax-H3-Image $0.02179.

• 2K text-to-image — $0.0504 vs $0.02333.

• 4K text-to-image — $0.113 vs $0.06879, or $0.03439 under M​iniMax's limited-time 50% discount.

• Edit with one reference — G​oogle does not publish a per-image edit price at these tiers; M​iniMax charges $0.02654 at 1K, $0.02658 at 2K and $0.07312 at 4K.

• Edit with three references — $0.03563 at 1K, $0.04177 at 2K, $0.07914 at 4K on M​iniMax.

Two details in that ladder are easy to miss and change the arithmetic. First, M​iniMax's discount applies only when the delivered image is at least 16,600,000 pixels, and it expires on 17 November 2026 at 16:00 UTC — six days after the weights went up. Price a pipeline on the discounted 4K number and you are pricing on a promotion. Second, M​iniMax's ladder is unusually flat: sixteen times the pixels for a little over three times the money, because the meter is tokens and the token count does not scale with area the way the sticker price suggests. G​oogle's ladder is closer to proportional. At 4K, the resolution the open-weight model was built for, that flatness is where the real saving lives.

Where a routing layer earns its place here is in not having to care which of those two meters you are on. OrcaRouter passes provider list price through with 0% markup, which means a vendor price change — including a promotion like M​iniMax's, and including its expiry — is live on our side the same day rather than waiting on a billing migration. Neither model in this comparison is on our catalogue: MiniMax-H3-Image is a week old with no hosted route, and Nano Banana 2.1 is reached through G​oogle's own API. What we do carry from that lineage is the previous generation, google/gemini-3.1-flash-image-preview, alongside O​penAI's image routes and a video route for M​iniMax's H3. The reason to say so plainly is that a piece comparing an unmeasured open-weight checkpoint to a scored closed one should not pretend the new checkpoint is one keypress away.

Only one side of this has a scoreboard

Artificial Analysis scores Nano Banana 2.1 at Elo 1160 on its text-to-image board, fourth place, over 6,055 samples, and Elo 1137 on its image-editing board, also fourth, over 5,968 samples. Both placements carry a rank range of fourth to sixth depending on the interval you read. Those are independent, sampled, and comparable to every other model on the board — including the ones you are already using.

MiniMax-H3-Image has no entry on either board. Not a low one; none. The vendor's card calls it “a leading open-weight image model” and the repository's showcase grid is the only evidence of output quality that exists, which is exactly the kind of evidence that is designed to look good. There is no vendor evaluation table either — no GenEval, no DPG-Bench, no human preference numbers, nothing to compare against a competitor's published evaluation. One month from now that may have changed. Today, the honest summary is that the cheaper model is cheaper and the measured model is measured, and the two facts are not in tension because they are about different things.

There is a second-order cost that the per-image price hides entirely, and it is the largest number in this article: 68.65 GB of diffusion transformer and VAE in BF16 and FP32, plus a text encoder that is not in the repository at all. MiniMax-H3-Image loads text_encoder/ from the MiniMax-H3 video model, 66.7 GB of Qwen3-VL, so a first local generation is a download of roughly 135 GB across two repositories. Add the GPUs: M​iniMax's own serving guidance for vLLM-Omni names H200, RTX PRO 6000 Blackwell and AMD Instinct MI350X, and multi-GPU sequence parallelism — --ulysses-degree on SGLang, --usp on vLLM-Omni — is the documented way to get acceptable latency at full resolution. At $0.02179 a 1K image, the API is cheaper than the electricity for a very long time. The open weights are valuable for control and for the licence's terms, not for unit economics at small scale.

Limits that decide more deployments than quality does

Resolution ceiling is the obvious one and it is not close. Nano Banana 2.1 stops at 4K. MiniMax-H3-Image goes to 20 megapixels through the API and roughly 16 megapixels in the card's examples — 4096×4096 at 1:1, 6272×2688 at 21:9, 5472×3072 at 4K 16:9. If your deliverable is a print-size plate or a background that will be cropped hard, the closed model is out of the running before quality enters the discussion, and the two models in this article stop competing at all.

Reference handling is the reverse. G​oogle's 14 references are typed — 10 objects, 4 characters — and the typing is doing work: it tells the model what role each image plays. M​iniMax's 16 are untyped slots addressed by position, which is more headroom and less structure. If your workflow is “put this person in this jacket against this backdrop”, G​oogle's typing is the feature. If it is “here are sixteen frames, hold the identity across all of them”, M​iniMax's ceiling is the feature. Neither is a superset.

Prompt budget is the third. Nano Banana 2.1 has 131,072 tokens of input context to work with. MiniMax-H3-Image accepts up to 32,000 characters on the served API but only roughly the first 4,096 tokens take effect with prompt enhancement off, and the local SGLang path truncates at 4,096 tokens outright. For a model whose pitch is instruction following, that is a tight budget, which is why M​iniMax shipped a companion: MiniMax-H3-Image-PE, a 27B FP4 fine-tune of Qwen3.8-27B that rewrites a short request into a long caption plus an aspect-ratio recommendation and emits it as JSON against a schema. It fits in about 19 GiB of VRAM. Its edit-side half is described as prompt-only and still in development, so the expander currently helps text-to-image and not much else.

And the licence, which has no counterpart on G​oogle's side because there is nothing to license. M​iniMax's community agreement requires separate prior written authorisation above one million US dollars of annualised revenue or if you offer hosted services, requires the words “MiniMax-H3-Image” displayed prominently on any commercial interface that uses it, restricts use to a defined territory, and forbids using outputs to train or improve other models. G​oogle's terms are a standard API contract. “Open weights” here means a licence you have to read, not a licence you can ignore.

What the two bills look like for a real month

Say you generate 20,000 images a month at 2K. Nano Banana 2.1 costs $1,008 at list, or $504 on the batch tier if the job can queue. MiniMax-H3-Image costs $466.60 at list through M​iniMax's API — or, self-hosted, whatever two to four of those GPUs cost you for the hours involved, which for most teams is more. Say instead you generate 2,000 images a month at 4K and the discount window has closed: $226 for G​oogle, $137.58 for M​iniMax. The saving is real and it is three figures a month, not five. That is the correct size of the prize, and it should be weighed against the cost of being the first person to find out where an unmeasured model breaks.

That asymmetry argues for a specific pattern rather than a winner. Route the production path to the model with the scoreboard, and send a slice of traffic to the cheap one to build your own evidence — a fixed prompt set, your own review, a fortnight of comparison. This is the case failover is actually for: an unfamiliar model treated as a candidate route rather than a replacement, with the known-good one behind it when a job comes back wrong. Six-step unguided sampling at 16 megapixels is also a long request by image standards, and it is not idempotent on M​iniMax's side — a client timeout is charged, and the O​penAI SDKs retry twice by default unless you set max_retries=0. Long, billed-on-timeout requests are exactly the ones you want a second route under.

The question that resolves this in about a month

Neither model is the wrong choice; the choice depends on whether you can afford to be the one measuring. If the deliverable needs a scoreboard before it needs a decision, Nano Banana 2.1 is the only side of this comparison that has one, and at 4K it costs $0.113 against M​iniMax's $0.06879 — a difference you will spend more than once arguing about internally. If the deliverable needs 16 megapixels, typed references do not matter and resolution does, and MiniMax-H3-Image is the only side that ships it.

What would settle it is the thing that has not happened: an independent placement. MiniMax-H3-Image is one day old in public, its licence terms were still being asked about in the repository's discussion thread hours after the weights appeared, and its own prompt expander is half-finished. When it turns up on Artificial Analysis with a sample count, the price gap stops being a proxy argument and becomes an actual trade. Until then, buy the measured model for the work that has to ship, and treat the cheap one as a candidate you are evaluating — not as a replacement you have decided on.