Hy4 matches Opus on Tencent's own agent test at 1/8 the price, but weights stay under lock
The cost gap is real: for a fraction of Opus's price, Hy4's agent scores are competitive on Tencent's own test and third parties confirm the cost advantage on Agent Arena. But until weights land, the "open-source" claim only holds via the API, and those headline numbers come from self-organized benchmarks, so the actual capability for self-hosters remains unverified.
Hunyuan Hy4 ships with the same 770B total / 49B active parameters as the August preview, but now covers visual understanding and Computer Use, including training on professional tools like Blender and CAD. Tencent's official APEX-Agents v1.1 score is 75.2, just above Claude Opus 5.5's 74.8, with output priced at $2.501 per million tokens versus Opus's $20. On Terminal-Bench 4.0, however, Hy4 scores 44.4 against Opus and Kimi K3.1 Pro's ~66, a 20-point gap.
Tencent's own internal evaluations for office scenarios (83.18 average, 96.15% pass rate) and finance (3.04 blind-test score) are self-organized, not public benchmarks. Third-party Agent Arena ranks Hy4 14th—up from the preview's 25th—with a per-task cost of $0.55, about 60% of Kimi K3.1 Pro's $1.46 and a score within a point.
The API went live on Tencent Cloud and OpenRouter the same day, but model weights are only promised for "the coming days". The preview was open-weight on day one; the official release is API-first. Third-party hands-on tests found most tasks done efficiently but not flawlessly, and community feedback on the preview noted noticeable long-thinking latency.
Tencent's own benchmark numbers should be read as marketing for its strongest home turf, not impartial evidence; third-party ranking at 14th is a more sober check.
The real divider is the open-source cadence: the preview was open-weight immediately, while the official release is API-first, making every capability claim provisional until weights appear.
Cost-performance rivalry is becoming the deciding factor among Chinese open-source models, as Hy4, Kimi K3.1 Pro, GLM 5.3, and DeepSeek trade wins across benchmarks.
The weight-vs-open-source distinction matters more than ever: open weights do not mean a model is open source, and the community is already calling that out.
With a 20-point gap on Terminal-Bench, Hy4 is not closing the gap with closed frontier models on serious coding; its strengths lie in cheaper agent tasks and specific business workflows.
Tencent's bet is on its own ecosystem: WorkBuddy, CodeBuddy, Yuanbao, and WeChat Work integration, which is where enterprise value concentrates even if independent developers just want a cheap API.