Tencent Hy4 launches as a model to test, not migrate to yet
Hy4 brings a cheaper API, million-token context and broader agent capabilities, but production teams should verify completed-task cost and reliability before migrating.
What launched
Tencent released Hy4 on November 4 with a one-million-token context window, text and image input, tool calling, structured output and new computer-use capabilities. Its API is live, with launch pricing discounted by 50% through December 31.
Why it matters
Long context combined with lower pricing could change the economics of agents working across large repositories, document sets or extended tool chains. The useful metric is not token price, however; it is the cost of a reliably completed task.
What we would do
We would not replace an incumbent model on vendor benchmarks and launch claims. We would run Hy4 in shadow traffic against real tasks, using the same tools, permissions and success criteria as the production model.
Hy4 is interesting enough to enter a serious production trial now, but not proven enough to become anyone's default model. The launch discount, one-million-token context, image input, tool calling and structured output make it a credible candidate for agent workflows that currently spend frontier-model money on long jobs. But Tencent says model weights and system cards are still coming, while broad independent benchmark coverage remains thin. For us, that changes the decision from wait to test, not from test to migrate. We would place Hy4 behind a multi-provider abstraction, send it shadow copies of real production tasks and measure cost per successful completion, latency, tool errors and recovery cost. If it wins only on document-heavy or visual work, we would route only those jobs. If it cannot remain reliable through long tool chains, the temporary token discount is not enough reason to accept new operational risk and vendor lock-in.
Turn insight into a working system.
We build multi-provider agent pipelines that shadow-test Hy4 against your current model, score completed tasks, and keep billing, permissions and audit outside any one vendor. If your production agent could benefit from lower-cost long-context reasoning, talk to us.