Tencent Hy4 Explained: Vision, Computer Use, Pricing, and the Vendor Numbers to Verify

What Is Tencent Hy4?

Tencent launched Hy4 on November 4, 2026, a new artificial intelligence model that adds two capabilities many business users have been waiting for: visual understanding and computer use. In plain terms, Hy4 can look at images and screens, and it can operate software through both command-line instructions and graphical interfaces. That combination matters because most real work does not happen in a chat window alone. It happens across documents, dashboards, browsers, and desktop tools, and a model that can only handle text will always need a human to bridge the gap.

Hy4 is available through Tencent Cloud, OpenRouter, and Tencent products including CodeBuddy and WorkBuddy. That distribution is important for two reasons. First, it means developers and businesses can reach the model through a cloud API or through tools they may already use, rather than waiting for a separate product launch. Second, availability through OpenRouter puts Hy4 side by side with models from other labs, which makes it easier to compare pricing and performance in the same workflow.

There is one point to be precise about from the start. Tencent’s announcement describes the model weights as a future release commitment, not a downloadable final checkpoint. In other words, you can use Hy4 through the listed services today, but you should not equate that access with the ability to download and run the model locally. That distinction matters for teams planning self-hosted deployments, and it is worth checking before you build anything that assumes local weights.

Hy4 Pricing: What It Costs to Use

Tencent kept Hy4’s published API pricing unchanged from the preview, which is good news for anyone who tested the preview and is now planning production use. The announced rates are $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens. Those are the company’s published figures, and they place Hy4 in the affordable range for a capable model, particularly for workloads where a large share of input can be cached and reused.

Pricing alone never tells the whole story, because the real cost of a model depends on how much output it generates, how often you repeat similar prompts, and how much work it can complete without human correction. Cached input pricing helps when you send the same documents or instructions repeatedly, which is common in finance, support, and document workflows. Output pricing matters most for long reports, code generation, and multi-step agent tasks, where the model writes far more than you do. Before committing, run your own representative workload through the API and compare the actual bill against your current model, because benchmark scores and price lists do not capture retries, corrections, or failed runs.

What Hy4 Can Actually Do: Vision and Computer Use

The headline capabilities of Hy4 are vision and computer use, and both deserve a plain explanation. Visual understanding means the model can interpret images, screenshots, charts, and documents that contain visual information, not just plain text. That is essential for finance, where data lives in tables and charts, for support, where problems arrive as screenshots, and for any workflow where a human would normally look at a screen before acting.

Computer use means Hy4 can combine command-line operation with graphical-interface operation. A model that can only use a command line is powerful for developers but useless for the many business tools that only exist as visual software. A model that can only click buttons is flexible but slow for bulk work. Hy4’s claim is that it can do both, switching between typed commands and on-screen actions as the task requires. If that works reliably in practice, it opens up agent workflows that previously needed custom integrations for every single application.

As with every new model launch, the gap between announcement and everyday reliability is where the real test happens. Vision and computer-use features are notoriously easy to demo and hard to make dependable across messy, real-world software. The sensible approach is to pilot Hy4 on a narrow, well-defined task with clear success criteria, measure how often it completes the task without intervention, and only then expand to wider use.

Hy4’s Reported Results: Read Them With Care

Tencent reported two sets of numbers with Hy4, and both need to be read with the right caveats. In the company’s evaluation with 213 finance professionals across 579 financial tasks, Hy4 scored 3.04, compared with 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3. Those are vendor-reported results, not independent validation, and the task set, scoring method, and judge details are Tencent’s own. The ranking is interesting, especially the lead over other efficient models, but it is not proof that Hy4 will outperform those models on your specific workload.

The second number is more striking and more fragile. Tencent reported that its internal hallucination rate fell from 15.76 percent for the Hy4 preview to 5.45 percent for Hy4. That is a large improvement, and if it holds up in independent testing, it would make Hy4 considerably more trustworthy for finance and business work, where a confident wrong answer is far worse than no answer at all. However, the announcement does not provide enough sampling and scoring detail to reproduce that rate, so treat it as a claim to verify, not a fact to rely on. The right response is to test hallucination on your own data, with your own checks, before trusting Hy4 with anything that affects money, customers, or compliance.

Hy4 vs the Other Models From the Biggest AI Week of 2026

Hy4 did not arrive in a vacuum. It landed just days after one of the most concentrated bursts of frontier model releases of the year, a week in which four major labs shipped significant models within four days. Understanding Hy4 means understanding what it is competing against, so here is how the week’s releases compare on the facts that have been publicly reported.

OpenAI shipped GPT-6.1 Terra, a mid-range model priced at roughly half the cost of GPT-5.6 Terra, and it is now generally available on Amazon Bedrock. OpenAI reported that Terra scored 15 percentage points higher than GPT-5.6 Terra on its Agents’ Last Exam evaluation, performed on par with GPT-6 Astra, and completed tasks more than three times faster than both. Those are OpenAI’s figures, not an independent test, and Terra takes up to 1 million input tokens with text and image input. Its input and output token prices are at least 50 percent lower than GPT-6.1 Sol and GPT-5.6 Terra, which positions it as an everyday-workload model rather than a flagship.

Google answered with Gemini 4 Carbon, announced on October 28, two weeks after Gemini 4 Argon. Carbon ranks number one on the v5.0 Artificial Analysis Intelligence Index with a score of 59, according to the research digest covering the week. DeepMind also released Genie 4 in the same week, a new generation of interactive world simulation with real-time, user-controllable worlds, which points in a different direction entirely: models that simulate environments rather than just answering questions.

Mistral released the weights of Large 4 on October 29, a 1 trillion-parameter natively multimodal model with 52 billion active parameters, keeping its promise to open the model after a preview. That release is philosophically different from Hy4’s approach: Mistral gives you the weights to run yourself, while Tencent is offering API access now and weights later. Neither approach is automatically better. Open weights suit teams that need full control, local deployment, or deep customization. API access suits teams that want to start immediately without infrastructure work.

The pattern across the week is clear. The frontier is no longer just about a higher score on a single benchmark. It is about efficiency, with Terra priced well below its siblings, about openness, with Mistral releasing weights, about new capabilities, with Genie simulating worlds, and about practical business use, with Hy4 combining vision and computer operation. For buyers, that means the right question is no longer “which model is smartest” but “which model fits the job, the budget, and the deployment constraints you actually have.”

Who Should Use Hy4, and For What?

Hy4 makes the most sense for teams whose work is visual, repetitive, and spread across multiple applications. Finance teams are the obvious first audience, given Tencent’s own evaluation, but the same logic applies to operations, support, and back-office work where staff constantly read screens, copy data between systems, and follow multi-step procedures. If your workflow involves reading a dashboard, checking a document, and then acting in another tool, a model with vision and computer use is built for exactly that chain.

Developers should look at Hy4 through CodeBuddy and WorkBuddy, where the model can work inside coding and productivity tools rather than as a standalone chatbot. The combination of command-line and graphical operation is particularly useful here, because real development work mixes terminal commands, code editors, browsers, and issue trackers. A model that can only handle one of those will always leave part of the job to you.

Small businesses and individual users should be more cautious. Hy4’s pricing is affordable, but computer-use agents need careful setup, permissions, and supervision before they touch real systems. Start with read-only tasks, such as summarizing documents or extracting data from screens, before granting any agent the ability to change files, send messages, or move money. The technology is exciting, but the guardrails matter more than the demo.

The Risks and Limits You Should Not Ignore

Every new model launch comes with risks that the announcement does not emphasize, and Hy4 is no exception. First, the performance numbers are vendor-reported. Until independent evaluators test Hy4 on public benchmarks and real business tasks, the finance score and the hallucination improvement should be treated as marketing claims with some supporting detail, not as established facts.

Second, computer-use agents introduce a new category of risk. A text model can give you a wrong answer, and you can ignore it. An agent that operates software can take wrong actions: clicking the wrong button, editing the wrong record, or sending the wrong message. That means permissions, approval steps, and audit logs are not optional extras. They are the core of any safe deployment, and they should be in place before the agent touches production systems.

Third, the weights question matters for long-term planning. If your strategy depends on running the model yourself, Hy4 is not ready for that today, because Tencent has only committed to a future weights release. Build your timeline around API access, and treat local deployment as a future option rather than a current capability. Teams that need open weights now should look at Mistral Large 4 or other openly released models instead.

Finally, there is the usual caution about pricing. Published per-token prices are only part of the cost. Agent workflows generate many more tokens than single-turn chats, because the model reads screens, reasons about actions, and retries when something fails. Budget for that multiplier, monitor usage closely in the first weeks, and set spending limits before you scale up.

How to Try Hy4 Safely This Week

If you want to evaluate Hy4 without taking on unnecessary risk, follow a short, disciplined pilot. First, pick one narrow task with a clear definition of success, such as extracting data from a set of invoices or summarizing a batch of reports. Second, run the same task through Hy4 and through your current model, and compare accuracy, cost, and time, not just the headline score. Third, keep a human in the loop for every action that changes data or contacts a customer, and log everything the agent does so you can audit mistakes.

Fourth, test the hallucination question directly. Give Hy4 documents with known facts, including some tricky cases where the correct answer is “this information is not in the document,” and see how often it invents an answer anyway. That single test will tell you more about whether Hy4 is safe for your work than any vendor benchmark. Fifth, check the pricing on your real workload, including cached versus uncached input, and compare the total bill against the value of the time saved. Only expand the pilot when all five checks look good.

Frequently Asked Questions About Tencent Hy4

What is Tencent Hy4?

Hy4 is Tencent’s AI model launched on November 4, 2026, adding visual understanding and combined command-line and graphical-interface computer use, available through Tencent Cloud, OpenRouter, CodeBuddy, and WorkBuddy.

How much does Hy4 cost?

Tencent’s published API pricing, unchanged from the preview, is $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached input tokens.

Can I download and run Hy4 locally?

Not yet. Tencent describes the model weights as a future release commitment, not a downloadable final checkpoint, so current access is through the listed APIs and products rather than local deployment.

How accurate is Hy4, according to Tencent?

In Tencent’s evaluation with 213 finance professionals across 579 financial tasks, Hy4 scored 3.04 versus 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3. Tencent also reported its internal hallucination rate fell from 15.76 percent for the preview to 5.45 percent. Both figures are vendor-reported and have not been independently validated.

How does Hy4 compare with GPT-6.1 Terra and Gemini 4 Carbon?

They target different strengths: Hy4 emphasizes vision and computer use for business workflows, GPT-6.1 Terra emphasizes everyday-workload efficiency at roughly half the cost of GPT-5.6 Terra, and Gemini 4 Carbon ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. The best choice depends on your task, budget, and deployment needs rather than any single ranking.

Is Hy4 safe to use for finance work?

Hy4 is promising for finance workflows, but its reported accuracy gains are vendor-reported and its hallucination improvement has not been independently reproduced. Pilot it on narrow tasks with human review and audit logs before trusting it with money, customers, or compliance decisions.