Hy Image 3.5 Preview: Tencent's $0.024 Image Model, Sourced

§ 01What Tencent shipped#

Two Tencent documents describe the model, and they do not agree on its limits. The announcement is the smaller sheet; the international API guide, stamped 11:26 the same morning, is the larger one.

The gap between five and twenty references, and between 2K and 4K, is probably the difference between a launch note written for consumers and a reference written for developers. Both are Tencent’s, so this page carries both. The first reply under the X post asked whether the model edits as well as generates; the API guide answers yes, with editing context carried across turns.

§ 02The price#

Tencent Cloud TokenHub bills images as tokens at a flat 10 RMB per million, then quotes a reference price per image. That makes the ladder exact.

A 2K image costs 25% less than a 3.0 image did, and a 4K image costs what a 3.0 image did. On the international API the announcement quotes $0.024 per image for the 2K tier; the international 4K price is not stated. Tencent charges for output images only: reference inputs and failed generations are free, which the X post put as “your reference images are free”. Against the two Western APIs most buyers will compare it with, at 1K resolution and each vendor’s own list price:

Nano Banana 2 costs about 2.8 times Tencent’s price at 1K and Nano Banana Pro about 5.6 times; at 2K Google’s list moves to $0.101 while Tencent’s stays at $0.024. GPT Image 2.5 straddles it, cheaper at medium and about 2.2 times dearer at high. Price is the one axis where every number on this page is a list price; quality is the axis where only Tencent has spoken.

01List price per 1K image, each vendor's own page; Tencent second from the bottom

02Two Tencent documents, one morning, two sets of limits; the API guide is the wider

03A major version about every two months since February, preview first, full release after

1 / 3

←→

§ 03The quality claim, and who made it#

Tencent’s announcement says the model was blind-tested by several hundred of its professional designers in a GSB format against Hy Image 3.0 and came out 30% better; the X post compresses this to “+30% win rate in human eval vs Hy Image3.0”. The same paragraph places 3.5 preview level with Seedream 5.0 Pro and slightly ahead of Nano Banana Pro and Qwen-Image-3.0 Pro. Three things to hold beside that. The evaluators are Tencent employees and the prompt set, comparison count, win-tie-loss split and margin are unpublished. No independent leaderboard had a 3.5 entry on launch day; the Arena text-to-image board scores the previous model, hunyuan-image-3.0, at 1151 on 173,099 votes, well below the frontier. And the usage number Tencent offers as its real test, image requests in Yuanbao up more than 50% during a two-month quiet rollout, measures appetite, not quality. It is a plausible launch claim from a lab that has published honest numbers before; it is still a vendor’s claim until someone else runs the prompts.

§ 04What is closed#

Hy Image 3.0 was the largest open-weight image model when Tencent released it: a 64-expert mixture of experts with 80B total parameters and 13B active per token, weights on Hugging Face, technical report on arXiv on September 28, 2025. For 3.5 preview, Tencent has published nothing about the model: no size, no architecture, no report, no weights plan. Tencent’s own open-model index still lists HunyuanImage-3.0-Instruct as the newest image entry. The two most-read questions under the launch post were “Did you forget open source” and “is this going open weight?”. Neither had an answer from Tencent when we read the thread at 11:46 UTC. The preview label leaves room: Tencent says Hunyuan has shipped a major version about every two months since it rebuilt its infrastructure in February, preview first and a full release after, and Hy4 preview followed that shape on August 28. Whether the image line follows Hy4 into open weights is the open question this record tracks.

§ 05Where you can use it#

The X post names only OnSolo and Miora for the free window; the announcement adds WorkRally. This is a model shipped into eight products on day one, most of them Tencent’s own. One reply under the post from an international user complained that the Chinese AI Studio now requires a mainland phone number and that earlier generations were lost; the international API is the route for everyone else.

§ 06Where CellCog stands#

Our stake, declared: CellCog runs an image agent on GPT Image 2.5, Flare by default and Sunburst where precision matters, and we chose it on quality with the price as a known cost. Every figure on this page was made by that agent. We do not route to Hy Image 3.5 and had no access to it before this record was written; we have no view on its output yet. What we can read is the price table, and at $0.024 for a 2K image Tencent has set a floor that the high-quality tiers of the models we use do not meet. If Tencent’s parity claim against Seedream 5.0 Pro survives an independent test, a model at this price gets tested on our infrastructure and this page records the result, in either direction.

§ 07What we are watching for#

• An independent score. A 3.5 preview entry on the Arena text-to-image leaderboard, or any third-party blind test with a published prompt set.
• The model card. Parameter count, architecture, a technical report, or a statement on weights. Any of the four flips the closed rows in the table above.
• The full release. Tencent’s two-month cadence and preview-first pattern suggest a non-preview 3.5; its price and whether it keeps the 0.15 RMB floor.
• The international 4K price and whether the announcement’s five-reference limit or the API guide’s twenty is the one that ships to consumers.
• Our own test. If we run it against our reference set, the numbers land here.

§ 08Update log#

This is a living page; when the story moves, the update lands here.

As of September 22, 2026, 12:00 UTC: page opened. The Tencent Cloud announcement and the TokenHub pricing table read in a browser; the international API guide read at 11:52 UTC (page stamp 2026-09-22 11:26:56); the X post read in a browser at 11:46 UTC and clocked from its id (02:40:42 UTC); Google’s and OpenAI’s pricing pages read at 11:52 UTC; the Hy Image 3.0 repository and the Arena leaderboard read at 11:50 UTC.

§ 09Sources#

• Tencent Cloud, Hy Image 3.5 preview release announcement (Chinese), Tencent Cloud official account, September 22, 2026, 10:35 CST.
• Tencent Hunyuan on X, launch post, 02:40:42 UTC, September 22, 2026; time computed from the post id.
• Tencent Cloud, Hy Image Generation API Guide, international edition, page stamp 2026-09-22 11:26:56.
• Tencent Cloud, TokenHub model pricing (Chinese), read September 22, 2026.
• Tencent Hunyuan, open-source model index, read September 22, 2026.
• Tencent Hunyuan, HunyuanImage-3.0 repository, README read September 22, 2026.
• Arena, text-to-image leaderboard, read September 22, 2026.
• Google, Gemini Developer API pricing, read September 22, 2026.
• OpenAI, API pricing, read September 22, 2026; GPT Image 2.5 token counts per image measured for our GPT Image 2.5 record, September 8, 2026.
• Our records: Qwen-Image-2.1 (September 20, 2026), What Is Tencent Hy4? (August 28, 2026).