Tencent ships Hy Image 3.5 Preview with multi-turn editing at $0.024
The preview accepts text and reference images, keeps editing context across turns, and undercuts Hy Image 3.0's listed API price by 25%.
By Ryan Merket · Published
Primary source: Tencent Hunyuan on X
Why it matters
Tencent is pairing conversational image editing with a lower per-image price, shifting the competition from one-shot generation toward iterative creative workflows that developers can embed in production software.
Tencent released Hy Image 3.5 Preview on September 22nd, adding conversational image editing and reference-image input to its Hunyuan generation API while pricing output at $0.024 per image.
https://x.com/TencentHunyuan/status/2102226552310419473
In its launch post on X, Tencent said the model handles text-to-image and image-to-image requests and produces images at up to 2K resolution. Tencent also claimed a 30% higher win rate than Hy Image 3.0 in human evaluations, though that figure comes from Tencent's own testing rather than an independent benchmark.
The larger change is in how developers interact with the model. Hy Image 3.5 Preview uses a messages-based interface resembling a multimodal chat API, allowing an application to submit text, reference images and prior editing turns as a single conversation. That moves Tencent's image product beyond isolated prompt-and-output calls toward an editing workflow that can retain context across repeated revisions.
Tencent's TokenHub documentation says the API supports text generation, reference-based generation and multi-turn editing. A request can carry as many as 20 reference images, each up to 20MB, while the input context extends to 100,000 tokens. Developers can also fix a seed to reproduce an output, select dimensions, add a custom watermark footnote or enable external-search enhancement.
The documentation exposes an important distinction from Tencent's shorter launch description. The standard generate_max_pixels setting tops out at a 2K area tier, matching the limit advertised in the X post. Developers can request images as large as 4096 by 4096 pixels by passing explicit dimensions through the size parameter. Width or height can range from 256 to 8192 pixels, provided the total image area remains within the API's 16,777,216-pixel ceiling.
That makes Hy Image 3.5 Preview a 4K-capable API under a specific request path, even though Tencent is marketing the general service around 2K output. The distinction matters for applications that need print assets, product imagery or other high-resolution files and would otherwise read the launch post as a hard 2K limit.
Tencent turns image editing into a conversation
Hy Image 3.0 used a conventional request structure centered on a prompt, dimensions and optional reference images. Hy Image 3.5 Preview instead accepts an ordered message history and returns an assembled_history object that developers can feed into the next request. Tencent handles the model's generated image and tool-call records inside that history, reducing the amount of state an application must reconstruct between edits.
The API responds synchronously, so developers receive the generated image without creating a separate job and polling for completion. Output links remain active for 12 hours, according to Tencent's documentation, which means production applications still need to copy generated files into their own storage.
The service can also route repeated requests carrying the same session identifier to the same inference instance. Tencent says this is intended to improve cache reuse during multi-turn editing. That feature points to the commercial target: applications where a user iterates on the same asset, rather than one-shot consumer image generation.
Tencent is charging less for that expanded workflow. Hy Image 3.5 Preview's stated $0.024 price is 25% below the $0.032 per-image reference price Tencent Cloud lists for Hy Image 3.0. The lower unit price gives developers a reason to test the preview despite the operational risk attached to integrating a model whose name and behavior can still change before a stable release.
The pricing also shows where Tencent is applying pressure. Image models increasingly compete on editing consistency, text rendering and the cost of repeated revisions, not simply the quality of a single generated frame. A multi-turn session can consume several images before a user accepts the result, making per-output pricing central to whether these tools can support high-volume creative software.
Tencent's 30% human-evaluation claim does not settle the quality comparison. Internal preference tests can establish that Tencent's evaluators favored the new version under Tencent's test conditions. Developers will still need to test prompt adherence, character consistency, typography and reference-image preservation on their own workloads. At $0.024 per attempt, Tencent is making that evaluation relatively inexpensive.