On November 16, 2026, Tencent announced Hy4 flash, a new AI model for everyday work. Standard API pricing is low, at 0.5 yuan per million input tokens and 2 yuan per million output tokens. In Tencent's internal evaluation, the cost per task came to one-twentieth that of its higher-tier model, Hy4. Low-cost models can be used for tasks such as creating presentations and processing documents. However, the actual cost of real-world work depends not only on API unit prices but also on the number of output tokens and the human effort needed to revise results.

AD

Hy4 flash targets everyday work, with 12 billion active parameters

Hy4 flash has 224 billion total parameters, of which 64 billion are allocated to Engram. The number of parameters active during inference is 12 billion. It supports multimodal input, including images, and can process a context length of up to 1 million tokens at once.

Tencent lists in-depth research, document processing and PowerPoint creation among its main use cases. The model can also build simple front-end screens and operate a PC. For complex, long-running software development, however, Tencent recommends Hy4, the higher-tier model it released on November 4. The company's approach is to use Hy4 flash for routine work and Hy4 for complex development.

According to the official announcement in Chinese, standard API pricing is 0.5 yuan per million input tokens, 2 yuan per million output tokens, and 0.02 yuan for cached input. The English version lists these as roughly $0.070, $0.281 and $0.003, respectively. The actual bill depends on the number of tokens used in each category: standard input, cached input and output.

At the time of the announcement, the model is available through Tencent Cloud TokenHub and WorkBuddy, as well as Yuanbao, WeCom, QQ, ima and Marvis. Support for OpenRouter is planned for the near future. Through November 30, it is free to use in WorkBuddy, and TokenHub applies a 50% discount to standard pricing. When estimating the cost of continued use, these limited-time offers need to be distinguished from standard pricing.

Average output tokens halved, and inference made faster

Tencent says it optimized everything from the model architecture to the training method and inference process to reduce costs. Pretraining uses hybrid attention, LatentMoE and Engram. The company says the size of the KV cache, which holds past information during generation, has been reduced to one-twenty-fourth that of Hy3 and one-sixth that of Hy4.

In post-training, the model inherited capabilities from Hy4 through self-distillation, followed by reinforcement learning. According to Tencent, this reinforcement learning cut the average number of output tokens generated across tasks by 50%. Even if the per-token output price is the same, generating fewer tokens to finish a task lowers the actual usage fee. Reducing output volume is therefore a cost factor separate from API unit prices.

For inference, Tencent combined megakernel, 4-bit quantization and DFly, and says it achieved a decoding speed of more than 150 tokens per second. Decoding refers to the process in which a model generates a response sequentially. However, the official announcement does not give details of the measurement conditions, such as the GPU used, batch size or input length. Judging response times in a real-world environment will require individual measurements.

AD

How API unit prices differ from per-task costs

Tencent evaluated not only token prices but also the cost of completing a single task. WorkIndex, an in-house benchmark, is built from representative tasks handled in WorkBuddy and incorporates human evaluation. Tencent says it emphasizes output quality, the length of text generated in a single response and the time required to complete a task.

WorkIndex consists of eight types of work. Creating, editing and co-authoring PowerPoint files are treated as separate items, and Markdown co-authoring and document editing are also included. Spreadsheet repair, audio and video processing, and visual understanding are also evaluated. Tencent also presented results from public agent benchmarks, saying that in overall effectiveness Hy4 flash slightly outperformed Hy4 Preview.

Tencent says Hy4 flash's cost per task on WorkIndex was one-twentieth that of Hy4, one-thirteenth that of Hy4 Preview and one-eighth that of DeepSeek-V4.1-Flash-1017. The DeepSeek comparison uses peak-hour pricing at the Max tier. However, these are the results of Tencent's internal evaluation, and the announcement does not disclose the task composition, pass criteria, handling of retries or a detailed cost breakdown.

Based on Tencent's official announcement in English and the Chinese version, the differences between the metrics can be summarized as follows. API unit prices are rates per million tokens, the 50% output reduction refers to the average number of output tokens across tasks, and the one-twentieth cost refers to the cost per task on Tencent's internal WorkIndex.

Metric Published figure / comparison Basis and conditions of comparison What to check when adopting
Standard API pricing Input 0.5 yuan, output 2 yuan, cached input 0.02 yuan Per million tokens each. Standard pricing announced November 16 Input and output usage, volume of cache applied, actual charges
Average output volume Reduced by 50% through reinforcement learning Average output tokens across tasks. A change resulting from training Output volume and quality of deliverables in actual work
Internal per-task cost One-twentieth of Hy4, one-thirteenth of Hy4 Preview, one-eighth of DeepSeek-V4.1-Flash-1017 Tencent's internal WorkIndex. DeepSeek comparison uses Max-tier peak-hour pricing Match with your own tasks, pass criteria, handling of retries
Actual cost of work Evaluated by each adopting company Until one deliverable of the required quality is completed Success rate, number of retries, API fees, human revision time

The 50% average reduction in output tokens and the one-twentieth per-task cost relative to Hy4 were measured under different conditions. Tencent has not disclosed how much each of the model configuration changes, training improvements and inference optimizations contributed to the cost reduction. In addition, API fees calculated from token usage need to be evaluated separately from the overall cost of work required to produce a finished deliverable.

Tencent also separately ran a blind evaluation with 85 internal WorkBuddy users. The test covered 27 types of tasks, 259 in total, and was conducted with model names hidden. According to the company, user satisfaction with Hy4 flash was close to that of Claude Haiku 5.5, and it had the lowest per-task cost and the shortest task completion time among the models compared. However, the evaluators were internal users, and neither specific satisfaction rates nor human revision times have been published.

Companies considering adoption could have both Hy4 and Hy4 flash perform the same document-editing or slide-creation tasks and compare the results against common pass criteria. If they also record the number of retries and the time people spend on revisions, along with API usage, they can judge how much the low API unit price actually reduces real work costs. Tencent also plans to open-source Hy4 flash and provide continuous updates. By comparing operating costs at standard pricing with output quality, companies can consider dividing the work: leaving everyday tasks to Hy4 flash and using Hy4 for complex development.