Tencent Hy4 Beat GPT in OSWorld and Learned to Control a Computer

⚡ The 5-Second Rundown

• Tencent has released Hy4 — the first model in the Hy lineup with visual input and computer control.
• It works two ways: terminal commands and GUI actions. Context window: 1 million tokens.
• In OSWorld 2.1, Hy4 scored 78.9 vs GPT's 75.2, but trailed Claude (81.8). In ProgramBench and Terminal Bench 4.0, the gap to both is noticeably larger.

🔍 What Happened

Tencent has released Hy4 — the first model in the Hy lineup that accepts not only text but also images, and can control a computer. It can operate both through the terminal and through a graphical interface. Its context window is 1 million tokens. The company itself describes its purpose as: development, workflow automation, document handling, and interface creation. This was reported by the BotHub blog on Habr.

The mechanics for the computer-use class of agents are familiar: the model receives the screen state, decides what to do next, and outputs either a shell command or a click in the interface. The numbers from the same post: ProgramBench — 30 for Hy4, 46 for GPT, 56.5 for Claude. Terminal Bench 4.0 — 44.4 / 55.5 / 66.4. OSWorld 2.1, where the AI performs tasks in a real OS — 78.9 / 75.2 / 81.8. In other words, in exactly one test out of three, Hy4 is ahead of GPT, and only by 3.7 points.

🧠 Our Take

Betting on computer agents is currently the most obvious way to enter the conversation about "a model that actually works." Chatbots with 1M context no longer sell, but "click through the admin panel for me" is still a live pain point. Our assessment: Hy4 is not an attempt to overtake Claude in general intelligence, but a claim on a niche where there is no established leader yet. One benchmark win out of three confirms this better than it refutes it.

What changes in practice. The combination of "visual input + a million tokens" hits long scenarios: parsing a batch of documents, going through a multi-step form without losing context along the way. For developers, the terminal part is more interesting — there it's 44.4 vs Claude's 66.4, and that's an honest signal that in complex shell tasks Hy4 is not yet a competitor. Documents and interfaces — yes; hands-on debugging — not quite.

Who wins: Tencent gets a seat in the conversation about agents rather than chat; teams that need a cheap routine-automation tool get a potential alternative. Who loses: those who built their positioning on "we're the only ones who click on the screen." Market experience shows that leads in such niches last months, not years.

⚠️ What We Don't Know

The main weak point is the source. This is a retelling in the blog of a systems integrator, not an official Tencent announcement with a tech doc. We don't know the model size, the license, the price, whether the weights are open, or the latency when working with the screen. All three benchmarks are self-reported measurements, not independent runs. What's worth checking yourself: API availability and how the model behaves on multi-step tasks, not on three representative sets.

Prediction: within a quarter, at least one independent OSWorld 2.1 run involving Hy4 will appear — on someone else's testbed or in an open leaderboard. If 78.9 holds up, the conversation about "caught up with GPT" will become substantive; if it drops, the debate will end on its own.

Disclaimer: the benchmark figures and capability descriptions come from a retelling in the BotHub blog; we have not verified the official Tencent announcement — a single source.