Tencent Hy4 explained

API cost calculator

Hy4 bills three token counts separately. Cached input costs one twentieth of uncached input, so a long session that re-reads the same documents costs much less per run than the headline input price suggests. Enter your own numbers, or start from a preset.

Can I use it today?

Choose what you need. The answer follows the status Tencent published on November 4, 2026, not a guess about what might arrive next.

What Tencent Hy4 is

Tencent Hy4 is the full version of the model Tencent previewed on August 28, 2026. The preview was text only. The full release adds two capabilities for the first time: vision and computer use. Tencent says it also scaled up pre-training and post-training over the preview, with gains in instruction following, long-context understanding, long-horizon tasks, and deep research.

Vision means the model reads images, not only text. Tencent lists fine-grained image processing, complex chart analysis, and mathematical visual reasoning. One demo locates a roof fitting about 2 metres across in a 2.8-billion-pixel aerial photo. Another generates an interactive 3D model. Both demos come from Tencent, and neither has an independent check attached.

Computer use means the model works through command-line tools and on-screen apps, and can combine the two in one task. Tencent says it trained Hy4 on professional software such as Blender and CAD tools, and on long tasks that span several applications. That training is aimed at real work: building a 3D scene, editing a document, or running a multi-step job across apps, rather than answering a single question.

Where Tencent aims it

Office work is the first use case in the announcement. On WorkBuddy, Tencent's work assistant, Hy4 is wired into chats, email, documents, calendars, and meetings through WeCom. It tracks project history, flags meetings that need preparation, and watches approaching deadlines. A Projects feature for team collaboration is described as upcoming, not as shipped.

Finance is the second. Tencent built thousands of workspaces and tens of thousands of tasks from real financial work, covering fund research, securities analysis, and quantitative investment. The blind evaluation in the benchmarks section below comes from that work. Frontend and game development are third: Tencent reports tuning for complex business logic, state management, and full-stack work, plus game work with its Interactive Entertainment Group across Unreal Engine, Unity, and Godot.

Science is fourth, and it is the one part with a download today. Tencent open-sourced two smaller models built with Hy4 and its Hyra 2.0 harness: HyCell for single-cell representations and HyPDE for physics fields. Those are separate models, not Hy4 itself. Hy4 is also described as moving into recursive self-improvement work, where the model helps build smaller models on the same architecture. That is a research direction, not a product you can use.

Specifications: what is confirmed and what is carried over

The full release announcement does not restate the parameter count, the active parameter count, or the context length. The figures in circulation come from the preview and from listings. This table keeps the two apart. Do not quote the preview figures as full-release specifications.

Preview figures come from the preview announcement and its Hugging Face page, as reported in coverage of the full release. The OpenRouter figures are listing figures, not Tencent announcement figures. The preview model card is tencent/Hy4-preview; the full model has no model card yet.

Benchmarks, with the conditions attached

Tencent publishes a 34-benchmark appendix with the full release. It pits Hy4 against the preview, GLM-5.3, DeepSeek-V4.1-Flash, and Kimi K3.1 Pro, plus two closed models, OpenAI GPT-6.1 Sol and Anthropic Claude Opus 5.5. Twenty of the 34 benchmarks are Tencent's own internal tests. Most comparison scores come from Tencent's own runs. The table below shows the rows developers ask about most. Every number is Tencent's own measurement. Read the notes under the table before you rank models on these numbers.

Read the table as a shape, not a ranking. Hy4 sits close to the closed models on computer use: 78.9 on OSWorld 2.1, against 81.8 for Claude Opus 5.5 and ahead of GPT-6.1 Sol at 75.2. It trails by a wide margin on terminal work: 44.4 on Terminal-Bench 4.0, against 66.4 and 66.8 for the two leaders. On ProgramBench it scores 30.0 against 56.5 for Claude Opus 5.5. Its one first place in this selection is APEX-Agents v1.1, 75.2 against 74.8, a gap of 0.4 points in Tencent's own run.

The internal tests show the gap Tencent admits. On Hy-FrontierMath T5, Hy4 scores 13.0 against 97.5 for GPT-6.1 Sol and 91.3 for Claude Opus 5.5. On ProofBench it scores 5.5 against 70.5 for GPT-6.1 Sol and 42.2 for Claude Opus 5.5. Those are Tencent's own hardest tests, and Tencent writes that a large gap remains to the top closed models on frontier math, long-horizon development, model development, and advanced visual questions. A reader who only sees the finance and OSWorld rows gets a different model from the one Tencent describes in full.

Gains over the preview

Tencent reports gains over the preview on its own suites, not on public leaderboards. The direction is useful. The size of each gain is not independently checked.

One note on the ProgramBench row: coverage of the appendix notes that Claude Opus 5.5 was blocked by content filters on 41 instances during Tencent's run. That condition is part of the comparison, and it is one more reason not to treat a single row as a settled ranking.

What the benchmarks do not prove

No independent reproduction was published as of November 5, 2026. Artificial Analysis, LMArena, and other independent leaderboards had no Hy4 score in the coverage checked for this page. Until one does, treat every score above as a vendor claim with a method note, not as a measured fact about your workload. The weights are the missing step: independent testers cannot run the full model until Tencent publishes them.

The safety claims need the same care. Tencent says it limited cybersecurity training data to defensive material, such as proof-of-concept generation and vulnerability repair, and removed exploitation, reverse engineering, and penetration testing material. Its Xuanwu Lab and Keen Security Lab audited the model and found no pronounced offensive risk in the areas they checked. On biosafety, Tencent reports 30.4 out of 100 on Hy-BioExperimentBench, its own test, below the high-risk threshold it set, with limited ability to turn biological knowledge into open-ended experimental designs. Those are Tencent's audits and Tencent's threshold. The promised system card is supposed to carry fuller safety results. It is not out yet.

Two different hallucination figures are reported, and they do not match. Tencent's English press release says the hallucination rate fell from 12.95 percent in the preview to 7.2 percent, with errors from unsupported assumptions down from 23.1 percent to 5.6 percent. Coverage of the Chinese release post reports a fall from 15.76 percent to 5.45 percent on the same kind of measure. The two pairs come from different posts and different evaluations. This page does not merge them into one figure. Treat both as reported, and wait for the system card to state which evaluation each pair belongs to.

Finally, do not confuse the preview with the release when you read older coverage. Pages written in August and September describe a text-only preview with open weights under Apache 2.0. Some still call Hy4 open source without qualification. The full release is API only today, with no license published and no date for the weights. Pages that present downloadable weights for the full model as a fact are reading the preview precedent as a guarantee. It is not one.

Pricing in full

The list price is unchanged from the preview. That is the one pricing fact all sources agree on. The discount is time limited, and the two discounts in play have different end dates. Do not budget on the discounted price past its end date.

Tencent Cloud TokenHub charges 50 percent off the list price through December 31, 2026. OpenRouter shows the same 50 percent discount on its Hy4 listing, at $0.417 input and $1.251 output, with Tencent Cloud as the only provider. OpenRouter does not state an end date for its discount in the coverage checked, so do not assume it matches the TokenHub date. WorkBuddy and CodeBuddy are free from November 4 to 15, 2026, with a limited free usage quota. After November 15, app use falls back to the API price.

Two labelling cautions matter for anyone quoting a price. First, Tencent's English press release rounds the list price to approximately $0.90 input, $0.045 cached input, and $2.70 output. Those rounded figures and the precise $0.834, $0.042, and $2.501 figures describe the same list price. The calculator on this page uses the precise figures. Second, the OpenRouter price is a gateway listing price, not a separate Tencent price. Gateway prices include the gateway's own terms and they move. Check the listing on the day you buy.

For comparison at the same batch size used in the calculator, 800,000 input tokens and 200,000 output tokens with no cache hits cost $1.17 at list price and $0.58 at the discounted price. With 70 percent of input served from cache, the same batch costs $0.72 at list price. Cache is where the saving sits, because cached input is priced at one twentieth of uncached input. A workload that re-reads the same repository in every run benefits far more than a workload that reads new context each time.

Budget warning: the 50 percent discount is a launch offer, not a list price. A project that only works at $0.417 input and $1.251 output depends on a price Tencent and OpenRouter can change. Run the calculator above at list price before you sign a customer up to a fixed fee.

How to connect it

On OpenRouter, Hy4 has the model id tencent/hy4. The listing accepts text and image inputs and returns text, with a 1,048,576 token context window and a 65,536 token completion limit. Tencent Cloud is the only listed provider, so there is no routing choice to make on OpenRouter today. The code below uses the OpenRouter OpenAI-compatible endpoint. Check the OpenRouter Hy4 page for current auth requirements before you run it in production.

OpenRouter, Python

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_OPENROUTER_API_KEY",
    base_url="https://openrouter.ai/api/v1",
)
response = client.chat.completions.create(
    model="tencent/hy4",
    messages=[{"role": "user", "content": "Hello!"}],
    max_tokens=1000,
)
print(response.choices[0].message.content)

OpenRouter, curl

curl -X POST https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer YOUR_OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "tencent/hy4",
    "messages": [{"role": "user", "content": "Hello!"}],
    "max_tokens": 1000
  }'

Keep your API key out of shared files and chat logs. A leaked key can be used to spend your quota. If a key leaks, create a new one and delete the old one. No official Tencent Cloud code sample was in the sources checked for this page, so this section uses the OpenRouter route only. For TokenHub, follow the Tencent Cloud docs for your account region.

One practical limit to plan around: the listing context is about 1 million tokens, but the completion limit is 65,536 tokens. A long output will stop at that limit. Build long jobs as a series of shorter calls, and check the listing again if a long output is cut short.

Preview to release: what changed, and what did not

April 2026: Tencent releases Hy3, a 295 billion parameter model, as the previous generation in the Hy family.

August 28, 2026: Hy4 preview launches as a text-only mixture-of-experts model with 770 billion total parameters, 49 billion active per token, and a context window of about 1 million tokens. Weights ship under Apache 2.0. Tencent notes excessive reasoning and over-verification as known issues in the early version.

November 4, 2026: the full Hy4 release adds vision and computer use, scales up pre-training and post-training, and publishes the 34-benchmark appendix. The model is live the same day through Tencent Cloud TokenHub, OpenRouter, and Tencent apps. Weights and a system card are promised in the coming days. Hy4 Flash is announced as coming soon. Two science models, HyCell and HyPDE, are open-sourced the same day.

What did not change matters as much as what did. The API list price is the same as the preview. The model is still reached only through hosted APIs and Tencent apps. The preview weights remain the only downloadable weights. If you tried the preview in August, the November release gives you vision, computer use, and a benchmark appendix to judge it by. It does not give you new weights, a license, or a system card.

Frequently asked questions

Sources

• Tencent, Tencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models (November 4, 2026): https://www.tencent.com/tencent-launches-hy4-joining-the-ranks-of-the-worlds-leading-open-source-ai-models/ (vision and computer use, WorkBuddy scores, finance evaluation, rounded pricing, free period and TokenHub discount, Hy4 Flash).
• OpenRouter, Hy4, API pricing and providers: https://openrouter.ai/tencent/hy4 (model id tencent/hy4, $0.834 and $2.501 list prices, 50 percent discount, one provider, 1,048,576 token context, 65,536 token completion limit, no EU in-region routing).
• Unite.AI, Tencent Releases Full Hy4 Model With Vision and Computer Use (November 4, 2026): https://www.unite.ai/tencent-releases-full-hy4-model-with-vision-and-computer-use/ (hosted access, preview contrast, 15.76 percent to 5.45 percent hallucination pair, safety checks, planned weights).
• MadRobot, Tencent Hy4 is out: it can use a computer but still trails Claude (November 4, 2026): https://madrobot.blog/2026/11/04/tencent-hy4-release-computer-use-open-weight/ (OSWorld, Terminal-Bench, and ProgramBench figures, gap admission, weights status, preview specifications).
• MACCUS, Tencent says Hy4 ranks among the best open models. Its weights aren't out yet (November 4, 2026): https://maccus.fr/en/tencent-hy4-ai-model/ (34-benchmark appendix, 20 internal tests, APEX-Agents and internal math figures, OpenRouter discount, Hugging Face status).
• ShiftDelete, Tencent Hy4: Open-Weight AI Model Adds Vision, Computer Use (November 4, 2026): https://shiftdelete.net/en/tencent-hy4 (vision and computer use, finance evaluation, WorkBuddy results, price and availability, HyCell and HyPDE).
• Reuters, Tencent rolls out AI model trained with limited cybersecurity data amid hacking fears (November 4, 2026): https://943jackfm.com/2026/11/04/tencent-rolls-out-ai-model-trained-with-limited-cybersecurity-data-amid-hacking-fears/ (defensive-only cybersecurity training, planned weights release).

This page is an independent explainer, not affiliated with Tencent. Figures that come from Tencent are labeled as reported by Tencent in the text. Prices were checked on November 5, 2026, and can change. The calculator runs in your browser. Nothing you type is sent to a server by this page.

More model guides

• Kimi K3.1 Pro, explained
• DeepSeek V4.1 Flash 1017, explained
• Mistral Large 4, explained
• LongCat 2.5, explained