Tencent Hy4: Vision, Agent Access and API Costs
Tencent Hy4 is the November 4, 2026 release with visual understanding and improvements to long-horizon agent work. Use WorkBuddy or CodeBuddy, or call the tencent/hy4 API route for text and image tasks. For computer-use workflows, pair the model with an environment that can supply observations, execute actions and check their results.
Content reviewed:
Use this model
API docs and console
- Open the Hy4 API listing
Provider account and usage billing apply.
Use cases
- Ask a focused question about a chart or screenshot, then check numbers against the source.
- Connect a tool-capable client to a bounded coding or document task.
- Budget every model turn and tool action in a longer workflow.
Hy4 official release versus Hy4 Preview
Choose tencent/hy4 for the November release. Hy4 Preview is a separate endpoint with released text checkpoints. Existing clients keep using Preview until you change the model selection, so check which model ID your client sends before testing image input.
OpenRouter’s November 4 listing identifies the official release and its first addition of visual understanding. The endpoint exposes image input, tool calls and structured output. Desktop actions additionally depend on the application’s tools, permissions and execution environment.
Hy4 API prices and launch-discount budgeting
On November 5, 2026, OpenRouter lists a 50% launch discount on the Tencent Cloud route: $0.417 per million input tokens, $1.251 per million output tokens and $0.021 per million cache reads. Listed regular rates are $0.834, $2.501 and $0.042 respectively. Tencent’s promotion runs through December 31, 2026; confirm the current provider quote before relying on it.
A 100,000-token input with 10,000 output tokens costs approximately $0.054 at the displayed reduced rate, or $0.108 at the listed regular rate, before tools and other charges. Tencent application subscriptions are separate from this API bill.
Long agent tasks can resend instructions, screenshots and tool results across many turns. Add actual billed usage for the complete run, including retries. Context capacity does not indicate how many tokens a task will consume.
Connect the exact Hy4 endpoint
The OpenRouter listing supports text and image inputs, text output, tool calling and structured responses, with a 1,048,576-token context and up to 65,536 completion tokens. The client and provider must support the particular feature you intend to use.
Create an account API key, configure the endpoint below and make a short text request first. For image input, use the provider’s chat image-input format. For computer use, connect an agent that captures observations and carries out approved actions; a text response alone does not manipulate a desktop.
import os
from openai import OpenAI
client = OpenAI(base_url="https://openrouter.ai/api/v1",
api_key=os.environ["OPENROUTER_API_KEY"])
response = client.chat.completions.create(
model="tencent/hy4", max_tokens=1024,
messages=[{"role": "user", "content": "List checks for reconciling an invoice with its purchase order."}],
)
print(response.choices[0].message.content)Start with a checkable visual task
For a chart, ask Hy4 to distinguish printed numbers from estimates inferred from the graphic. Supply a source table when exact calculations matter. For a screenshot, ask for visible controls and a proposed next action, then verify the target before execution.
For a document workflow, retain the original files and check quoted totals, dates and source locations in the generated report. A successful tool execution confirms that an action ran; it does not establish that every conclusion in the report is correct.
Can you self-host the November Hy4 release?
Use the hosted release for current work. Tencent said at launch that weights and a system card would follow; no public November-release weights are available yet. The existing Hy4 Preview files describe a separate model; they do not establish this release’s parameter count, memory requirement or serving recipe.
If self-hosting is required for the November release, no files or runtime instructions are available yet; do not size a deployment from Preview. For currently downloadable alternatives, compare Kimi K3.1 Pro and Ripple’s instruction-tuned releases, each with its own license and hardware requirements.
Before you use it
- Long context and tool support still require task budgets and checks on the final files, calculations and actions.
How to get started
Select the November Hy4 endpoint, start with a checkable text or image task, and record provider billing and complete task time.