Tencent Just Put a ¥0.15 Price Tag on AI Images. That’s Not a Typo.

Tencent Hy Image 3.5: RMB 0.15 per image, betting developers will do the math

On September 22, 2026, Tencent released Hy Image 3.5 Preview. On the same day, Alibaba held its 2026 Apsara Conference in Hangzhou, unveiling the latest progress of the Qwen series and plans to train trillion-parameter-level models. Tencent chose this timing to launch its new product; the intent to grab attention is written on the calendar.

This preemptive launch has a data backdrop. In SuperCLUE's August 2026 Chinese text-to-image leaderboard, Tencent's previous-generation model Hy-Image-3.0 scored 77.82, ranking 10th. ByteDance's Doubao Seedream 5.0 Pro scored 91.93, and Alibaba's Qwen-Image-3.0-Pro scored 91.60, ranking third and fourth. A 14-point gap is a full generation behind in this field.

Hy Image 3.5 Preview is Tencent's response.

Capability: Editing consistency is the focus, independent verification not yet in place

Hy Image 3.5 Preview supports text-to-image and image-to-image, can upload up to 5 reference images per request, supports multiple aspect ratios and up to 2K resolution, and has multi-turn dialogue creation capability, enabling continuous editing based on historical context.

The parameters are not stunning in the second half of 2026. The main claim is editing consistency: Tencent says the model can keep the subject from drifting after repeated edits and re-renders. Text rendering has also improved; the company says accuracy for generating Chinese and English characters has increased. These two capabilities point to commercial scenarios: posters, storyboards, and e-commerce images that require multiple revisions.

Tencent organized hundreds of professional designers for a GSB blind test internally, claiming Hy Image 3.5 Preview's overall win rate is 30% higher than Hy Image 3.0, on par with ByteDance Seedream 5.0 Pro, and slightly better than Google Nano-Banana Pro and Alibaba Qwen-Image-3.0 Pro.

These data come from Tencent's internal tests. As of publication, no independent evaluation institution has included Hy Image 3.5 in a leaderboard, and the 30% improvement has not been externally reproduced. Verification results from benchmarks such as SuperCLUE are the first variable to watch.

One grayscale data point is more direct than the blind test: Hy Image 3.5 Preview was in grayscale testing in the Yuanbao app for two months, during which image-generation requests increased by more than 50%. Whether users repeatedly use it says more than evaluation scores.

Price: RMB 0.15 per image, about one-ninth of GPT Image 2.5

Tencent Cloud API pricing: RMB 0.15 per 2K image, USD 0.024 per image internationally, charged only for output images, with failed images not billed.

For comparison: GPT Image 2.5 Sunburst, which ranks high on independent leaderboards, costs about USD 0.21 per image. Tencent's international pricing is about one-ninth of that. ByteDance Seedream 5.0 Pro's output price is USD 0.045 per image; Tencent is still nearly half cheaper.

This is not a promotional price. China's AI image generation market has a compound annual growth rate of over 30% and is expected to grow to USD 38.15 million by 2030. Tencent's logic: while the market is not yet settled, use extreme cost-effectiveness to buy developer call volume.

There is precedent for low pricing. Tencent's Hy3 language model is known for cost-effectiveness and topped OpenRouter's weekly ranking for several consecutive weeks after release. Tencent wants to replicate the same playbook in image generation.

But whether RMB 0.15 per image can cover inference costs depends on the pace of marginal cost decline after scaling. AI chip supply and computing cost fluctuations are the second variable to watch.

Ecosystem: Integrated into six apps, Yuanbao grayscale data is the hidden card

The model is not Tencent's differentiator. Deployment channels are. Hy Image 3.5 Preview is integrated into Yuanbao, WorkRally, OnSolo, Miora, WorkBuddy, and ima. Among them, Miora, WorkRally, and OnSolo offer limited-time two-week free trials.

WorkRally targets visual asset generation for film and television production, OnSolo serves series and game content creation, and Miora positions itself as a creative agent platform. Tencent lets the model validate within its own apps first, then opens the API to external developers.

ByteDance and Alibaba are taking another path. ByteDance Seedream 5.0 Pro emphasizes model design capabilities. Its official release proposed four major breakthroughs: complex information visualization, interactive precise editing, realistic image and portrait texture, and native multilingual input. ByteDance demonstrated the model's ability to simultaneously present a timeline, bar chart, pie chart, and research equipment photo in a single image.

Alibaba is betting on parameter scale. At the Apsara Conference, Alibaba CEO Wu Yongming announced that Qwen4 has entered training, with parameters expanding to 5 trillion to 10 trillion, and the next-generation video model will be released in November. Image generation is part of Alibaba's omnimodal strategy.

ByteDance bets on model capability. Alibaba bets on parameters and infrastructure. Tencent bets on app distribution and cost-effectiveness.

Trend: From "looks good" to "one image handles information"

In early 2026, 36Kr noted in an analysis that image generation models are shifting from "generating good-looking images" to "efficiently solving practical problems." In August, SuperCLUE added a "complex information layout" evaluation dimension, examining infographics, storyboards, and UI interface generation capabilities. The highest and lowest scores differed by 74.2 points, the most polarized of all dimensions.

Many models can generate good-looking images. Few can handle large amounts of text, tables, and flowcharts in a single image. In commercial scenarios, the latter is more valuable.

Tencent is aware of this. Hy Image 3.5 Preview's editing consistency and text rendering improvements target commercial scenarios. Some film and television studios report improved model performance in professional workflows. Some users point out that ByteDance Seedream 5.0 Pro remains stronger in complex information layout. For enterprises that need to produce professional infographics, whether a model can accurately handle data visualization may be more decisive than price.

What to watch

Whether Hy Image 3.5 Preview can help Tencent narrow the gap depends on three verifiable indicators:

First, independent evaluation results. Tencent claims parity with Seedream 5.0 Pro, but the conclusion comes only from internal blind tests. If independent benchmarks such as SuperCLUE show a significant gap, it will directly affect developer confidence.

Second, sustainability of low pricing. RMB 0.15 per image is effective in the customer acquisition phase, but maintaining it long-term requires inference costs to continue falling. Amid computing cost fluctuations, how long this price can hold depends on Tencent Cloud's subsequent pricing adjustments.

Third, conversion rate. After the two-week free period ends, how many enterprise users are willing to continuously call the API at RMB 0.15 per image is a harder indicator than integration count.

Tencent's iteration pace matters here. Since rebuilding its infrastructure in February 2026, Hunyuan large models have maintained a major version iteration every two months on average. The "preview first, formal version later" model is using speed to make up for a late start.

The competitive focus of China's AI image generation has shifted from "whose benchmark score is higher" to "who can solve more concrete business problems at lower cost." Tencent has placed its bet at RMB 0.15 per image. The next data point is paid call volume after the two-week free period ends.