Hunyuan Hy4 Official Release: A Cheaper, Multimodal Agent Model That Can Drive Another AI
Agentic coding models now have a strong open-source option that undercuts a competitor by 62% per task while matching its agent benchmark score, so teams running long unattended agent runs can cut costs significantly. Hy4's Computer Use also enables one harness to drive another, which opens a practical path to supervised multi-agent pipelines.
Hy4's official version expands pre-training and post-training, brings native multimodal and Computer Use, keeps API prices at preview levels, and plans open-source weights. Terminal-Bench 4.0 jumps from 26.3 to 44.4, WebCraftBench from 52.7 to 78.9, 3DArena-Modeling from 23.4 to 37.1, and Tencent's GameDev and GameLogic scores rise 19.1% and 9.8%. Agent Arena data puts Hy4's per-task cost at $0.55 versus $1.46 for Kimi K3.1 Pro, with Hy4 on the Pareto frontier of quality and cost.
Hands-on tests in WorkBuddy cover four scenarios: a one-click video explaining Ohm's law (built with a custom Skill for one-minute videos), a fully programmed Canvas 2D sheep-grazing game with day/night cycles and fireflies, a handwritten CSS/JS/SVG milk tea flowchart with parallel steps and branching, and an interactive ESP32 course with a rotatable 3D board and simulated current. The same harness then uses Computer Use to control a separate DeepSeek Harness, letting one agent supervise another to finish rendering the video. Hy4's conversational tone, with quips like "Coming right up, faster than delivery," shows a deliberate effort to reduce the sterile AI feel during long async jobs.
The benchmark jump from preview to official is unusually large, suggesting the preview was deliberately undercooked to stage a memorable launch.
The $0.55 per agentic task undercuts a direct competitor by 62% while landing on the Pareto frontier, which makes cost the decisive advantage in agent workloads where long runs multiply token usage.
The human-like dialogue quirks like "Coming right up, faster than delivery" are a UX tactic to keep users engaged during long unattended agent runs, but they also hint at a deeper trend: agent interfaces are trying to simulate co-worker presence.
Using one harness to control another via Computer Use turns a single-agent bottleneck into a supervisor-subordinate structure, which is a lightweight way to add oversight without building complex orchestration.