Hy Image3.5 vs Nano Banana 2: Two Ways to Buy a Picture

Nano Banana 2 — the consumer name for Gemini 3.1 Flash Image Preview — has been the default image model inside the vendor's app, Search's AI mode, Lens and Flow since it replaced Nano Banana Pro there, and it is the only model in this comparison series that you can call through OrcaRouter today. Tencent Hunyuan Hy Image3.5 preview arrived on 22 September 2026 with a rate card that looks, at first glance, almost identical: $0.024 per image from Tencent Cloud International against the vendor's $0.067 per 1K image, $0.101 at 2K and $0.151 at 4K. Two cheap image models, a few cents apart. The resemblance stops the moment you look at how each one bills and what each one has actually been measured at.

The interesting question here is not which is cheaper. It is what you are buying at each price point, and whether the one that costs more is buying you something you cannot get from the one that costs less.

Two rate cards, two billing philosophies

Hy Image3.5 is billed per delivered output image and nothing else. ¥0.15 per 2K image on Tencent Cloud in China, $0.024 per image internationally, with discarded and failed generations unbilled. There is no input charge to model, no token arithmetic to do, and no penalty for iterating. It does text-to-image and image-to-image in one model, keeps earlier turns as context across a multi-turn session, accepts up to five reference images per request, covers the standard aspect ratios, and caps output at 2K. Preview access runs through Yuanbao, ima, Miora, WorkRally, OnSolo and WorkBuddy plus Tencent Cloud TokenHub, MPS and VOD, with a free trial on Miora, WorkRally and OnSolo from 22 September to 7 October.

Gemini 3.1 Flash Image is billed the way a language model is billed, because that is what it is underneath. Input is $0.50 per million tokens and text output $3 per million, with image output metered separately at $60 per million — which works out to $0.045 for a 512-pixel image, $0.067 at 1K, $0.101 at 2K and $0.151 at 4K. Batch mode roughly halves those numbers. It is free at the point of use for Gemini app users, and it holds a 1M-token context window, so a request can carry a genuinely large amount of text alongside the image.

• Price per 1K image — Hy Image3.5 $24 equivalent at 2K; Gemini 3.1 Flash Image $67 at 1K and $101 at 2K

• Price at 4K — Hy Image3.5 does not offer it, the ceiling is 2K; Gemini 3.1 Flash Image $151 per 1,000 images

• Billing basis — Hy Image3.5 charges per delivered output image; Gemini 3.1 Flash Image charges per token, input and output separately

• Failed generations — Hy Image3.5 does not bill them; token-billed models charge for the tokens they spend getting there

• Reference images — Hy Image3.5 up to five per request; Gemini 3.1 Flash Image conditions on images within a 1M-token context window

• Independent placement — Gemini 3.1 Flash Image Elo 1,122 at #6 of the Artificial Analysis text-to-image board from 17,440 samples; Hy Image3.5 has no public arena entry

• Access — Gemini 3.1 Flash Image through Google's surfaces and OrcaRouter as google/gemini-3.1-flash-image-preview; Hy Image3.5 through Tencent's apps and cloud only

The sample count is the part that matters

Artificial Analysis has Gemini 3.1 Flash Image at Elo 1,122, sixth on the text-to-image board, from 17,440 samples — the largest vote count of any model in the top ten, collected in February 2026 and stable since. On the editing board it sits at 1,106 and ninth, from 14,171 samples. A number backed by seventeen thousand votes is not a marketing figure. It is the closest thing this category has to a settled result, and it puts the model within reach of MAI-Image-2.6 and Grok Imagine Image 2.0 while costing a fraction of what either charges.

Hy Image3.5's evidence base is a single vendor-run blind preference study. Tencent says hundreds of professional designers inside the company compared it blind against its predecessor and against the field, producing roughly a 30% improvement over Hy Image3.0, parity with Seedream 5.0 Pro, and a slight edge over Nano-Banana Pro and Qwen-Image-3.0 Pro. Every one of those figures is vendor-reported and unreproduced. Notably, the claim is a marginal edge over Nano-Banana Pro — the previous generation of Google's image model, not this one. Tencent does not claim a win over Nano Banana 2 anywhere in its launch material, and no third party has run the comparison.

This is the one matchup you can run on a single key

Hy Image3.5 is not on OrcaRouter. We host no Tencent image model — the only Tencent entry in the catalogue is the text-only Hy3 line — so getting to it means Tencent's own cloud and several third-party platforms. Nano Banana 2 is a different story: it is live in our catalogue as google/gemini-3.1-flash-image-preview, one OpenAI-compatible endpoint, provider list price passed through at 0% markup so a Google price change is live here the same day.

That makes this the only comparison in the series where half the question is already answered for you. Point your existing client at the same base URL you use for every other model, call the Gemini image model, and you have a production-grade baseline at a documented $67 per thousand 1K images with seventeen thousand votes behind it. Then run the same prompt set through Tencent's preview — free on three of the applications until 7 October — and compare cost per image you would actually ship. Because OrcaRouter also carries openai/gpt-image-2, google/imagen-4.0-ultra-generate-001 and grok/grok-imagine-image behind that same endpoint, the baseline you test against is not limited to one vendor, and automatic failover means a degraded route does not take your pipeline down with it.

Bottom line

Buy Nano Banana 2 if you want the settled answer. It costs more per image than Hy Image3.5 at every resolution the two share, but it has a 1M-token context for conditioning, a 4K tier that Tencent's preview does not offer, seventeen thousand independent votes placing it sixth in the world, and it is reachable through one endpoint you are probably already using. There is no version of this decision that requires you to take a vendor's word for anything.

Buy Hy Image3.5 if the economics are the point. At $0.024 per 2K image with no charge for discarded attempts, a workflow that iterates heavily is cheaper on Tencent's model by a margin that a per-token rate card cannot match — and the five-reference-image limit and multi-turn context make it a legitimate editing tool, not a toy. What you give up is the 4K tier and every scrap of independent verification. Tencent's own blind study is the only performance evidence that exists, and it does not claim a win over the model you would be replacing. Treat that as the honest state of play, and the test is worth running before the free trial window closes on 7 October.

OrcaRouter puts 200+ models behind one OpenAI-compatible endpoint at provider list price with 0% markup, so a vendor price change is live the same day.