Tencent Hunyuan released the official version of Hy4 on November 4: 770B total parameters, an agent benchmark score of 75.2, and an API output price roughly one-eighth that of Claude Opus 5.5. Tencent says it slightly outperforms Claude Opus 5.5. Our view is that model competition is shifting from parameter count to “the cost of completing a task,” although the claim that it is the “strongest open model” cannot be verified until the weights are actually released.
What this is
Hy4 and the preview released in August both have 770B total parameters and 49B active parameters. The upgrade comes primarily from additional training and post-training. Computer Use refers to a mechanism that enables the model to operate command lines or graphical interfaces to complete tasks. For the first time, the mechanism adds image understanding and computer operation to Hunyuan’s flagship model, with support for professional software such as Blender and CAD.
Tencent has connected the API to Tencent Cloud TokenHub and OpenRouter, pricing output at RMB 2.501 per million tokens. Public information on combined input and output pricing remains inconsistent, so the price comparison here mainly refers to output pricing. In Tencent’s official benchmark, Hy4 scored 75.2 on APEX-Agents v1.1, while Claude Opus 5.5 scored 74.8.
Its score on Terminal-Bench 4.0, however, was only 44.4, about 20 points behind Opus 5.5. The third-party Agent Arena ranking placed it 14th, also putting substantial distance between its results and Tencent’s claim that it has “joined the ranks of the strongest open models.” As of publication, the official weights had not been released; Tencent said only that they would become available within days.
Industry view
The most important development is the combination of price and work-related capabilities. Hy4’s official tests show that it can handle computer operation, office collaboration, and financial tasks at a cost significantly below that of leading closed models. Third-party tests have also found that it can perform tasks such as price comparison, converting screenshots into web pages, and interactive website creation. Although its performance remains unstable on some complex tasks, it is better viewed as a productivity-oriented model than a universally leading general-purpose model.
Our view is that Tencent’s advantage lies not only in the model, but also in its owned distribution channels, including WeCom, Tencent Meeting, and WorkBuddy. When enterprises deploy agents, permissions, data boundaries, and connections to existing software often matter more than a single benchmark score. Tencent may be looking to embed the model directly into existing workflows rather than sell only an API.
The risks are equally clear. The office and financial results came from evaluations organized by Tencent, and external expert blind testing is not equivalent to a public, reproducible benchmark. “Open weights” also do not automatically mean fully open source, because training code, data licenses, and reproduction conditions have not been clarified in parallel. If the official weights remain delayed, developers will face a higher cost of independently verifying the model’s capabilities.
Impact on regular people
For enterprise IT: Lower-cost models make it easier for enterprises to experiment with agent automation. Before deployment, however, teams still need to verify whether data leaves the corporate network, whether call logs can be audited, and whether permissions for computer operation are tightly restricted.
For individual professionals: Tasks such as organizing materials, comparing prices, and building websites may be handed off to models more quickly. At the same time, businesses will redefine which work can be automated, shifting more job value toward judgment, communication, and accountability.
For the consumer market: Tencent’s own office assistants will be the first to benefit from the model’s lower price, and related features may reach mainstream software sooner. Whether third-party applications can sustain low prices in the long term will depend on model call volumes, the duration of subsidies, and future price changes.