Tencent released Hy4 Flash on November 16, a model that charges about US$0.070 for a million tokens read and US$0.281 for a million written, and says internal users rated it close to Claude Haiku 5.5 in a blind test. The pitch is cheap, quick work on office-style tasks rather than frontier reasoning, which is where most enterprise token spend actually lands.
Hy4 Flash is the newest member of the Hy family, which Tencent announced on its corporate site. It has 224 billion total parameters, 64 billion of them Engram parameters, but activates only 12 billion per token. It accepts multimodal inputs and a context window of up to 1M tokens. Cached input tokens cost about US$0.003 per million.
Tencent credits several engineering choices for the price. Hybrid attention, LatentMoE, and Engram features shrink the KV cache to 1/24 the size of Hy3’s and 1/6 the size of Hy4’s. Reinforcement learning cut average end-to-end output tokens by 50 percent, which lowers the bill on every task. Tencent also reports decoding above 150 tokens per second, helped by megakernel optimization, DFly, and 4-bit quantization.
The cost comparisons rest on WorkIndex, a benchmark Tencent built from typical tasks in WorkBuddy, its workplace assistant. Its tasks span slides, documents, broken spreadsheets, audio and video handling, and image understanding. On that measure Tencent says Hy4 Flash costs 1/13 as much per task as Hy4 Preview, 1/20 as much as Hy4, and 1/8 as much as DeepSeek-V4.1-Flash-1017 at the Max tier with peak-hour pricing. Tencent also says it slightly beats the larger Hy4 Preview, which has 770 billion total and 49 billion active parameters, on overall effectiveness.
The human comparison is the headline claim. In a blind evaluation, 85 internal Tencent WorkBuddy users worked through 259 tasks across 27 categories. Tencent says Hy4 Flash reached satisfaction close to Claude Haiku 5.5, with the lowest per-task cost and the shortest completion time of all models tested. The company did not publish a satisfaction score in its announcement text, so “close” has no number attached here.
That matters because every input to this comparison is Tencent’s. The benchmark is internal, the testers are the company’s own employees using its own product, and the rival models were run under conditions Tencent chose. Tencent also lists a separate set of seven public agent evaluations, including Terminal-Bench 4.0 and OSWorld 2.1, but the announcement text gives no results for them. Independent testing has not yet confirmed any of it.
Availability is broad inside Tencent’s ecosystem: Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with OpenRouter integration listed as coming soon. Until November 30, WorkBuddy users get Hy4 Flash free, and TokenHub charges 50 percent off standard prices. Tencent also says it plans to open-source the model.
Tencent positions the lineup as a split. It recommends Hy4 Flash for research, document work, slide decks, basic front-end pages, and computer use, and points users to the larger Hy4 for long-horizon software engineering. Hy4 itself shipped on November 4, and Tencent says it ranks second among Chinese models on the Artificial Analysis Intelligence Index and top of the open-source field on Agent Arena.
For teams paying per token on document and slide workloads, the OpenRouter listing is the milestone to watch. Once it goes live, you can run Hy4 Flash and Haiku 5.5 on your own task set. Tencent published no cost ratio against Haiku, so that run is the only way to learn it, and the same test shows whether the 1/13 and 1/20 gaps over Tencent’s own Hy4 models hold outside WorkIndex.
Tencent, “Tencent Launches Hy4 flash to Power High-Speed, Cost-Efficient AI Deployment for Everyday Workflows,” tencent.com, November 16, 2026.