Hy Image3.5 Preview on OnSolo: The Character Has to Survive Episode Forty

The interesting claim in Tencent's launch of Hy Image3.5 preview is not the price, and it is not the benchmark. It is a production claim, and it is specific: a character designed once should hold across every episode of a series, and an edit should refine that character rather than restart it. That is the whole economics of short-drama production in one sentence. Overseas short drama is shot and shipped in volumes where a series is routinely sixty to a hundred episodes, each one needing a fresh set of stills for the character sheet, the keyframe pass and the episode thumbnails. If the face drifts between episode twelve and episode forty, the cost of fixing it is not one generation, it is a re-shoot of the asset library. Tencent put Hy Image3.5 preview live on 22 September 2026, and it is already inside OnSolo, the company's AI tool for overseas short-drama creation, whose members get a free two-week window that opened the same day.

What OnSolo actually is, and why the pairing is the story

OnSolo is not a general image playground. It is a one-stop overseas short-drama creation surface that runs the whole chain — an idea becomes a script, the script becomes character design, the character design becomes a storyboard, and the storyboard becomes a finished film. Every stage of that chain is a place where a model that cannot hold an identity costs the team money, and the first three stages are nothing but stills. Character sheets are stills. Storyboards are stills. Keyframes are stills. The video stages are downstream of them.

That is why the interesting part of this integration is not that a text-to-image model is available in a video tool. It is that the same model is doing the stills at every stage of the same chain, with the earlier turns carried forward. In a pipeline where the character sheet, the storyboard frame and the keyframe were previously produced by three different calls to three different models, each one re-derived the face from scratch and each one drifted a little. Multi-turn editing with retained context is the mechanism that stops that, and reference-image conditioning is what lets a director point at an approved sheet and say "this one, in this shot."

Tencent is leading with three improvements over Hy Image3.0 that map directly onto that workflow:

• Character consistency — the headline claim. A subject survives repeated edits and re-renders without drifting, which is the difference between a character sheet that is an asset and one that is a one-off.

• Text rendering — improved accuracy on in-image Chinese and English lettering. Short drama lives on titles, on-screen captions and episode cards, and text that garbles is text a human has to fix by hand.

• Single-element refinement — editing one element of a frame without disturbing the rest, which is what "an edit refines instead of restarting" means in practice. Change the coat, keep the face. Change the background, keep the light on the face.

The rate card, and why it is the second-most interesting number

The price is the part that will travel: ¥0.15 for a 2K image on Tencent Cloud domestically, US$0.024 an image on the international endpoint, and billing only on images the API actually returns. A request that fails or is filtered is not a charge. For a pipeline that iterates — generate, look, feed it back, generate again — that billing model matters more than the headline figure, because the discarded attempts are where a per-call rate card quietly doubles.

Put it next to what the independent board says the leaders cost:

• Hy Image3.5 preview — $0.024 per 2K image at Tencent's published international rate, billed on delivered output only.

• GPT Image 2.5 Sunburst (max) — about $0.21 per image, converted from OpenAI's rate as tracked at $210.72 per 1,000 images.

• Grok Imagine Image 2.0 — $60.00 per 1,000 images on the same board.

• MAI-Image-2.6 — $38.90 per 1,000 images.

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

The ceiling is 2K, the request takes up to five reference images, and both text-to-image and image-to-image run on the same endpoint under the identifier hy-image-v3.5-preview. One figure in circulation does not survive contact with Tencent's own announcement: a 4096×4096 output spec reported by one outlet. Tencent says "up to 2K" in as many words. Treat the 4K number as an error.

The free window, and what it is actually good for

Miora, WorkRally and OnSolo are running a free two-week window that opened 22 September and closes 7 October 2026. For a short-drama team the honest use of that window is not to admire the model. It is to run the one test that decides the purchase: take a character sheet you have already approved, hold it across forty consecutive edits, and look at frame forty next to frame one. That is the claim under test, and it is a test your own assets can settle in an afternoon. The window has a fixed expiry date, which is the point of it.

Where the trial ends and a pipeline begins, the useful discipline is to keep the generation call behind an interface you can repoint. OrcaRouter runs on a 0% markup model — the provider's list price is passed through unchanged, so a vendor price revision is live on our side the same day rather than waiting on a repricing cycle — and 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, so the model itself means Tencent's own cloud and the Tencent products above. What we provide is the stable, priced baseline you measure it against — openai/gpt-image-2, google/gemini-3.1-flash-image-preview, the Imagen 4 family and grok/grok-imagine-image all sit behind that one endpoint, which is what makes "run it against your own workload before the window closes" a practical instruction rather than an aspiration.

What has been measured, and by whom

Tencent's evidence is a GSD-style blind preference test, the pairwise human methodology Chinese labs use for this, 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 thirty percent, lands level with ByteDance's Seedream 5.0 Pro, and edges past Nano-Banana Pro and Alibaba's Qwen-Image-3.0-Pro. The methodology is a reasonable one. The participants are Tencent employees. The prompt set and the raw preferences have not been published as anything anyone can re-run. It is a vendor number, and we are reporting it as one.

Two other Tencent figures deserve the same label. The company says the model ran in a grayscale rollout inside Yuanbao, its consumer assistant, for about two months before the public announcement, and that image-generation requests in that product rose roughly fifty percent over the period. Both are the company's own telemetry, and both are the kind of number a launch post is entitled to lead with.

The independent picture is thinner than the launch. Artificial Analysis's text-to-image board carries no Hy Image3.5 row at all. What it carries from Tencent is the previous generation: HunyuanImage 3.0 Instruct at 964 Elo and HunyuanImage 3.0 at 945, ranked 65th and 70th on a 156-model board. 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, Grok Imagine Image 2.0 at 1154 — and the leader, GPT Image 2.5 Sunburst (max), is at 1197. A thirty percent improvement over a model near 945 does not obviously land at 1078, and it certainly does not reach 1197. That is not a refutation of the claim; it is the absence of the evidence that would settle it, which is exactly why the first third-party Elo for this model will be the most informative number published about it this quarter.

What to watch, and what to do before 7 October

Three things are unresolved and each has a date attached. The first is the independent measurement — nobody outside Tencent has run this model against a public prompt set, and until somebody does, the thirty percent is a claim rather than a result. The second is the release itself: this is a preview, and Tencent's stated cadence is preview first, official version later, which is the subject of a separate piece in this batch. The third is the free window, which closes on 7 October 2026 and is the only free look anyone outside Tencent has had at the model.

For a short-drama team the decision is unusually clean, because the feature that matters is testable with your own material and the price is not the obstacle. Take an approved character sheet, run it through the episode count your show actually has, and judge the drift with your own eyes. If the face holds, $0.024 for the stills layer of a hundred-episode series is close to a rounding error against the rest of the budget. If it does not hold, you have lost an afternoon. What you should not do is take the thirty percent on faith, because the one thing this launch does not contain is a measurement that came from outside the building it was made in.