Hy Image 3.5 API and Pricing, Explained Through a Real Pipeline

Hy Image 3.5 Preview is Tencent's image model, the latest in the line many people still search for as Hunyuan Image. It reached OpenRouter on October 5, 2026, as tencent/hy-image-v3.5-preview, and we added it to the Image Generation node and to chat the next day. One model handles text-to-image and image-to-image, it accepts up to 20 reference images in a request, and it renders at four resolution settings from 1K to 4K.

Its price is the part that needs explaining. The model is not listed per image. It is listed per output image token, which tells you almost nothing until you know how many tokens an image uses. We measured that when we onboarded it, so this post starts there.

The spec sheet

The thirteen ratios cover the shapes most work needs: 9:16 for Stories and Shorts, 4:5 for feed posts, 16:9 for video frames, and 4:1 for a website header, plus the in-between portrait and landscape shapes. A seed makes a result repeatable, so you can change one word of the prompt and compare like for like. The full capability list is on the Hy Image 3.5 model page.

What Hy Image 3.5 costs per image

The listed rate as of October 2026 is $0.0000016 per output image token, which is $1.60 per million image tokens. To turn that into a price per image you need the token count, and the provider does not publish one per resolution. So we rendered the same prompt at each setting on October 6, 2026, and read the usage the API returned.

Three things stand out.

1K, 1.5K and 2K cost the same. Every image up to 2K used the same 15,000 tokens, so a 2048-pixel square costs no more than a 1024-pixel one. If you are going to use the image at all, there is no saving in asking for less than 2K.

4K costs about a third more, not four times more. A 4096-pixel square has four times the pixels of a 2K one and used 20,000 tokens instead of 15,000. That is cheap for the size. Look at it before you commit, though: in our one 4K render, the soft out-of-focus background showed visible color banding, which a print or a tight crop would make obvious.

References and the prompt added nothing. A request with one reference image used the same 15,000 tokens as a plain prompt, and the prompt itself was not billed. We tested one reference, not twenty, so treat a heavy multi-reference request as likely the same price rather than proven.

Those token counts come from one afternoon of tests. A token-priced model can count differently for a different shape or a different provider build, which is why Treza quotes the measured amounts before a run and then bills what the provider actually charged.

Generate and edit in one model

With a prompt alone, Hy Image 3.5 generates. With a prompt plus references, it edits or composes. Attach one photo and describe the change, and you get that photo with the change applied. Attach a product, a character and a setting, and describe how they fit together, and it builds one image from all of them. The ceiling is 20 references, more than the 14 that Nano Banana 2.1 and Seedream 5.0 Flash accept.

We ran the single-reference case when we onboarded the model. The reference was a product still of a folded navy linen shirt, and the prompt asked for the same shirt on a white marble counter in a bright studio. It came back as the same shirt, same color and linen texture, collar and placket intact, now resting on veined marble in side light. That is the edit most product teams want: keep the item, change the set.

It was not fast. The reference edit took about 52 seconds, against 17 to 23 seconds for plain prompts at 1K and 2K and 44 seconds at 4K. In a pipeline that runs in the background that rarely matters, but if you are iterating by hand on references, expect to wait.

How it compares on the same node

All three of these models sit in the Image Generation node's model menu, so the comparison is practical, not theoretical.

The Hy Image and Nano Banana 2.1 figures are what we measured at onboarding, and the Seedream 5.0 Flash figure is its listed flat price, all as of October 2026. The Seedream 5.0 Flash pricing post has the detail on that model.

The pattern that follows: Seedream 5.0 Flash is still the cheapest per image, and Nano Banana 2.1 was the fastest in our tests. Hy Image 3.5 earns its place in two cases. The first is large output, because it is the least expensive way on the node to get a 4096-pixel square. The second is a composition with more than 14 inputs, a product line, a cast of characters, a set of brand references, where it is the only one of the three that takes them all in one request.

We are not going to rank how good the images look, because that depends on your prompt. The test is cheap: put the same prompt and references through two models on two nodes and compare.

What you actually pay on Treza

Hy Image 3.5 through a Treza pipeline is metered from a prepaid credit balance. Treza quotes the measured amounts above, plus its credit markup, before a run. One credit equals $0.01, packs start at $5, credits never expire, and there is no subscription. A failed generation charges nothing, and that includes a prompt the provider refuses.

Calling Hy Image 3.5 as an API

Direct access means provider credentials, learning the request shape, error handling, and your own retry logic before you have built anything.

Through a Treza pipeline, Hy Image 3.5 is a node on a canvas. Place the Image Generation node, pick the model, set the resolution and aspect ratio, wire references into its reference input, and publish the graph as a single versioned endpoint. Your app calls the typed /invoke route with JSON in and JSON out, or points an existing OpenAI SDK at the OpenAI-compatible route. If you later swap the node to a model from another lab, the endpoint your product calls stays the same.

The same pipeline is reachable from an agent. Treza's hosted MCP server exposes pipelines as tools that Claude and other MCP clients call directly. In chat, asking for Hy Image 3.5 by name and attaching references does the same thing with no canvas at all.

Putting it in a pipeline instead of a single call

A single image is rarely the deliverable. The more useful pattern is Hy Image 3.5 as the first node of several.

Compose a key frame from your references, then pass it to a video node as the first frame. The image to video generator page has that pairing set up, and the post on setting a first-frame image for AI video generation covers the handoff. A fuller AI video pipeline puts a script step ahead of the image and narration, captions, or publishing behind the video, then reruns the whole graph on a schedule or from an API call.

Starting from a blank canvas, the AI image generator page has the image side set up, and the AI video generator page puts a video model behind a single generate action. For thumbnails, the AI thumbnail generator can send its best frames through an Image Generation node as references to repaint them into one scene.

Frequently Asked Questions

How much does Hy Image 3.5 cost per image?

It is listed at $0.0000016 per output image token as of October 2026. In our tests on October 6, 2026, a 1K, 1.5K and 2K image each used 15,000 tokens, which is $0.024 an image, and a 4K image used 20,000 tokens, which is $0.032.

Is Hy Image 3.5 the same as Hunyuan Image?

It is Tencent's image model line that many people know as Hunyuan Image. On OpenRouter and on Treza it appears as Hy Image 3.5 Preview, with the model id tencent/hy-image-v3.5-preview.

Do reference images cost extra with Hy Image 3.5?

Not in our tests. A request with one reference image used the same 15,000 tokens as a plain prompt. The model accepts up to 20 references per request.

What resolution does Hy Image 3.5 render?

The node offers 1K, 1.5K, 2K and 4K. A square came back at 1024, 2048 and 4096 pixels for the 1K, 2K and 4K settings, and at 1536 pixels with no setting at all.

How long does Hy Image 3.5 take to render?

In our tests, 17 to 23 seconds at 1K and 2K, 44 seconds at 4K, and about 52 seconds for an edit with one reference image. That was a small sample from one day, so treat it as a rough guide.

Can I use Hy Image 3.5 in chat?

Yes. Ask for Hy Image 3.5 by name in chat, attach reference images if you have them, and the image comes back in the conversation.

Can I call Hy Image 3.5 through an API without integrating Tencent myself?

Yes. On Treza the model is a node inside a pipeline you publish as a versioned HTTP endpoint, with a typed /invoke route and an OpenAI-compatible route, so your product calls one API whichever image model sits behind the node.