Tencent releases Hy4 flash with cheaper tokens and a private-benchmark cost claim
Tencent’s new Hunyuan model charges 2 yuan per million output tokens and targets everyday office work. Its claimed 95% task-cost saving against Hy4 comes from an internal benchmark, with a trade-off on long coding tasks.
- Hy4 flash lists output at 2 yuan per million tokens, against 18 yuan for Hy4.
- Tencent claims a 95% reduction in task costs on its private WorkIndex benchmark.
- The company still recommends the larger Hy4 for long software-engineering tasks.
Tencent, the company behind Tencent Cloud and the Hunyuan family of AI models, released Hy4 flash on November 16 for research, documents, slides and other everyday work. The launch pairs sharply lower API prices with a more conditional proposition: Tencent says the model completes tasks at one-twentieth the cost of Hy4 in its own internal evaluation, according to the launch report.
For developers and office users, the consequential distinction is between cheaper tokens and cheaper completed work. Tencent’s published rates establish the former. The latter depends on how much output a task needs, the quality of the result and whether a smaller model can finish it reliably.

Tencent Cloud’s standard TokenHub rates, before temporary launch discounts, compare token charges rather than completed-task costs. Source: Tencent Cloud.
A cheaper tier for an established model family
Tencent introduced Hunyuan as its proprietary foundation model on September 7, 2023, opening it to enterprises in China through Tencent Cloud APIs. At that launch, Tencent said Hunyuan was already connected to 50 of its own products, including games, meeting software and document tools. Its role extends from developing models to distributing them through its cloud and applications.
It announced the larger Hy4 on November 4, 2026, introducing visual understanding and improvements to long-horizon agent tasks. Flash now adds a cheaper option within that generation.
Hy4 flash accepts multimodal input and supports a one-million-token context window. It has 224 billion total parameters, including 64 billion Engram parameters, but activates 12 billion per token, according to the . Total model size therefore should not be read as the amount of computation used for every token.
