Hot AI News

Tencent’s Hy4 appears on OpenRouter at half the preview’s price

Tencent's Hy4 is now on OpenRouter at $0.417 per million input tokens, about half the preview's price, but the listing does not say how it differs from August's preview.

Tencent’s Hy4 appears on OpenRouter at half the preview’s price
Image: OpenRouter

Tencent’s Hy4 model is now listed on OpenRouter, priced at $0.417 per million input tokens and $1.251 per million output tokens through Tencent Cloud, according to the official OpenRouter page. That is roughly half the rate of the Hy4 preview.

What the OpenRouter listing actually says

The page is mostly an API guide. It shows the model ID tencent/hy4, a text-only Chat Completions endpoint, and a single provider, Tencent Cloud. OpenRouter says it bills pay-as-you-go with no markup on the provider’s price. The page also mentions one further price item without listing it, so the full price table is only on the model page.

Benchmark chart for Tencent Hy4 preview from the model card, section 'A New Flagship Generation'
Image: Tencent

The supported features are streaming, tool calling through tools and tool_choice, structured output through response_format, and reasoning controls through reasoning and reasoning_effort. EU in-region routing is not currently available for this model.

The listing gives no context length, release date, parameter count or description. It also does not say whether Hy4 is a final release or the same model as the preview. Anything beyond price, endpoint and features therefore comes from earlier coverage of the preview, not from this page.

Benchmark chart for Tencent Hy4 preview from the model card, section 'A New Flagship Generation'
Image: Tencent
Advertisements

The preview it seems to follow

Tencent released a preview of Hy4 on August 28, 2026, as open weights, according to an explainer from eesel.ai that cites Tencent’s announcement and the Hugging Face model card. That card describes a mixture-of-experts model with 770 billion total parameters, about 49 billion active per token, and a 1 million token context window. The license is Apache 2.0. Weights are published in BF16 and FP8 builds on Hugging Face, GitHub, ModelScope and GitCode.

The same explainer lists the preview’s API price as $0.834 per million input tokens and $2.501 per million output tokens, under the model ID hy4-preview. A search summary of OpenRouter’s earlier listing, tencent/hy4-preview, gives about $0.83 and $2.50. The new tencent/hy4 entry charges about half of that on both sides.

Hy4 preview benchmark appendix chart
Image: Tencent

The explainer says the preview is free in Tencent’s WorkBuddy and CodeBuddy tools and also usable in Yuanbao and ima. These are Tencent’s offers and were not checked against the new listing.

Who this changes things for, in practice

For developers already routing through OpenRouter, the practical change is a cheaper Tencent option with a stable-looking ID. A job that sends 10 million input tokens and 2 million output tokens would cost about $6.67 at the listed rate. At the preview’s rate of $0.834 and $2.501 it would cost about $13.34.

Hy4 also gives open-weight users a hosted way to try a very large model without owning the hardware. The explainer cites an official vLLM recipe that puts local memory needs at roughly 1,848 GB for BF16 and about 924 GB for MXFP8. That puts self-hosting out of reach for most teams, so a metered API is the realistic route.

Quality claims remain Tencent’s own. According to the explainer, Tencent’s model card reports 92.3 on GPQA Diamond, 85.4 on Terminal Bench and 82.9 on SWE-bench Multilingual. The same source says Tencent’s internal blind evaluation scored Hy4 at 2.99 out of 4, against 2.94 for Kimi K3 and 2.92 for GLM-5.3. That test used 163 raters and 203 tasks and is not independent. The explainer also says Hy4 is not on the Artificial Analysis index, and that it ranks about fifth in LMArena’s Code Arena WebDev with a provisional “AutoEval” label.

The preview was also offered with caveats. The explainer says Tencent openly lists flaws such as over-long reasoning and over-verifying its own work. Teams that pay per output token should measure how much text the model produces on their own tasks before assuming the lower rate means a lower bill.

Advertisements

What is still missing

Several basics are absent from the listing. There is no context window, no statement on whether images are accepted (the page lists text only), and no explanation of why the price fell. It is also unclear whether tencent/hy4 replaces tencent/hy4-preview or runs alongside it, and whether Tencent has changed the model since August.

This is not the first Chinese model to reach OpenRouter at a low price. We reported on Kimi K3.1 Pro on OpenRouter last month, and Hy4 now joins that group of large models reachable through one key. Tencent has not published anything we could read alongside this listing, so the timing and the pricing logic remain unexplained.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Alex CarterAI News & Big Tech Correspondent

Alex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like