Hy Image3.5 Preview Is Live: Tencent Puts a 2K Image Model at $0.024 an Image

Tencent opened Hy Image3.5 preview to the public on September 22, 2026, and the figure that will travel furthest is not a benchmark. It is the price: ¥0.15 for a 2K image on Tencent Cloud, US$0.024 an image on the international endpoint, billed only on the images the API actually returns. Set that against the top of the independent leaderboard — GPT Image 2.5 Sunburst at roughly $0.21 an image — and Tencent is selling at about a ninth of the current market leader, while adding two capabilities its previous generation did not have in one product: image-to-image editing and multi-reference conditioning. The capability claim is larger still, and softer. Tencent says a blind human evaluation run inside the company put the new model 30% ahead of Hy Image3.0. Nobody outside Tencent has reproduced that, and as of today no independent board carries a Hy Image3.5 row at all.

What Tencent actually put on the API

Hy Image3.5 preview is a text-to-image and image-to-image model served through Tencent Cloud, listed under the model identifier hy-image-v3.5-preview and reachable through the company's TokenHub API surface. Both directions are live at launch, which matters more than it sounds: the previous generation's editing story was thinner, and "both T2I and I2I available" is the part of the announcement that changes what a pipeline can be built out of.

The specs Tencent is leading with, stated as the company states them:

• Resolution — output up to 2K, with a set of common aspect ratios rather than a single square.

• Reference images — multiple-image conditioning for identity and layout control, described across Tencent's Chinese product materials and most launch coverage as up to five images per request. At least one aggregator reported 20; five is the figure we can attribute, and we are not upgrading it.

• Modes — text-to-image, image-to-image, and multi-turn editing, where an edit is applied against the conversation's accumulated context rather than each request starting clean.

• Text rendering — Tencent claims improved accuracy on in-image Chinese and English lettering, which has been the standard weak point of the whole category.

• Composition — better handling of prompt-specified layout, and stronger photo-editing behaviour on real photographs as opposed to generated ones.

• Consistency — the headline word in the announcement. Tencent's claim is that a subject survives repeated edits and re-renders without drifting.

One number in circulation does not survive contact with the vendor's own post: a 4096×4096 output spec reported by one outlet. Tencent's own announcement says "up to 2K" in as many words, and every other source agrees. Treat the 4K figure as an error, not a roadmap.

The price is the whole announcement

Strip the adjectives out of the launch and what remains is a rate card. On Tencent Cloud the 2K image costs ¥0.15; on the international endpoint the same call is US$0.024. Billing is on output images only — a request that fails or is filtered is not a charge. That is a deliberate undercut of the entire Western field, and it lands hardest against the model currently sitting at number one on the independent text-to-image board.

Put the per-image numbers side by side, because they are the real comparison:

• Hy Image3.5 preview — $0.024 per 2K image at the published international rate.

• GPT Image 2.5 Sunburst (max) — about $0.21 per image, converted from OpenAI's $30 per million image-output tokens.

• Grok Imagine Image 2.0 — $0.06 per 2K image on xAI's direct API, or $60 per 1,000 images as tracked by Artificial Analysis.

• Microsoft MAI-Image-2.6 — $38.9 per 1,000 images on the same board, roughly $0.039 each.

• Seedream 5.0 Pro — $90 per 1,000 images, the model Tencent claims parity with.

The gap is not a discount, it is a different category of product decision. A team generating 50,000 images a month pays about $1,200 on Hy Image3.5 preview and about $10,500 on GPT Image 2.5 Sunburst. Whether the quality difference justifies that is exactly the question the missing independent benchmark fails to answer — and it is the reason the price is doing so much work in this launch.

Where a rate card becomes a routing decision is when you want the option without the migration. OrcaRouter runs on a 0% markup model: the provider's list price is passed through unchanged, so a vendor price cut — like Tencent's international rate here, or any future revision to it — is live on our side the same day rather than waiting on a repricing cycle. One API key covers 200-plus models, so a cheap generator and an expensive one can sit behind the same endpoint and be swapped by configuration instead of by rewrite. (To be precise about our own catalogue: we do not currently route Hy Image3.5 preview — the price comparison above is Tencent's published rate against other vendors' published rates, not a quote from us.)

Where the +30% comes from, and what it does not cover

The benchmark in the announcement deserves to be read carefully, because "30% better" is doing a lot of work with very little behind it in public.

What Tencent describes is a GSD-style blind comparison — the same methodology Chinese labs use for pairwise human preference — run with what the launch materials describe as several hundred of Tencent's own professional designers. On that evaluation, Hy Image3.5 preview beats Hy Image3.0 by roughly 30%, lands level with ByteDance's Seedream 5.0 Pro, and edges past Nano-Banana Pro and Alibaba's Qwen-Image-3.0 Pro. Three things are true at once: the methodology is a reasonable one, the participants are Tencent employees, and the results have not been published as a dataset anyone can re-run. It is a vendor number. We are reporting it as one.

The independent picture is a different shape. Artificial Analysis's text-to-image board, read on the day of this launch, carries no Hy Image3.5 entry and no Hy Image3.0 entry under that name. What it carries from Tencent is the prior generation: HunyuanImage 3.0 Instruct at 964 Elo and HunyuanImage 3.0 at 945, ranked 65th and 70th, with HunyuanImage 2.1 at 886 down at 99th. For scale, the models Tencent says it has caught sit far above that — Seedream 5.0 Pro at 1078, Qwen-Image-3.0-Pro at 1088 — and the board's leader, GPT Image 2.5 Sunburst (max), is at 1197.

That is not a contradiction of Tencent's claim; it is the absence of the evidence that would settle it. A 30% improvement over a model sitting at roughly 945 Elo does not obviously reach 1078, and it certainly does not reach 1197. Which is precisely why the claim needs an outside run rather than a dismissal, and why the first third-party Elo for Hy Image3.5 will be the most informative single number published about this model all quarter. Note also that the AA entries are the fal-hosted variants of the Hunyuan image family, so the comparison is between Tencent's model and Tencent's model as served by someone else — close enough to be useful, not identical enough to be a verdict.

It has been running inside Tencent's products for two months

The preview label understates how long this model has been in front of users. Tencent's materials describe a grayscale rollout inside Yuanbao — its consumer assistant — running for about two months before today, during which image-generation requests in that product rose roughly 50%. That is the kind of number a launch post is entitled to lead with and it is still a company's own telemetry.

What is harder to argue with is where the model is now reachable. Beyond the API, Tencent has put Hy Image3.5 preview into a list of first-party products: Yuanbao, the ima knowledge tool, and the creative and productivity surfaces Miora, WorkRally, OnSolo, and WorkBuddy. It is also wired into Tencent Cloud's media stack — Media Processing Service and Video on Demand — which means the intended buyers are not only developers building image features but teams already transcoding and editing video on Tencent infrastructure.

There is also a clock on the free access. Miora, WorkRally, and OnSolo are running a two-week free window that opened September 22 and closes October 7, 2026. If you want to form your own opinion of the model before committing to a pipeline, that window is the cheapest way to do it, and it expires on a fixed date.

The cadence is the real signal

Tencent rebuilt its image line in February 2026 and has shipped a numbered generation roughly every two months since, each one arriving first as a preview and only later as an official release. Hy Image3.5 preview is the current step in that sequence, not a one-off. Read as a schedule rather than a launch, the announcement says something more useful than any single benchmark: Tencent intends to keep landing image generations on a two-month clock, at a price point the rest of the field has not matched, and it is willing to ship a preview before the independent boards have caught up.

For a team deciding what to build on, that argues for treating Hy Image3.5 preview as a genuinely usable option with one unresolved risk — nobody outside Tencent has measured it — and structuring the decision so the risk is recoverable. Run it against the workload you actually have during the free window. Keep the generation call behind an interface you can repoint, which is the same discipline that lets you move between a $0.024 model and a $0.21 one without a sprint. And watch for the first independent Elo, because that number, not the 30%, is what will tell you whether Tencent has closed the gap it says it has closed — or merely priced around it.

The open question is narrow and specific: the price is already the best in the category, and the capability list is competitive on paper. What is missing is a measurement that did not originate inside the company that built the model. Until one exists, the honest reading of Hy Image3.5 preview is a strong price attached to a strong claim, with the claim still awaiting its first outside witness.