Tencent released Hy4, its new Hunyuan model, on 4 November, and its release page lists an API price of $0.834 per million input tokens and $2.501 per million output tokens. In Tencent’s own benchmark table, Hy4 scores 75.2 on APEX-Agents v1.1, an agent test, against 74.8 for Claude Opus 5.5 and 72.9 for DeepSeek V4.1 Flash-1017 in Tencent’s testing.
By our arithmetic, Hy4’s output price is one-eighth of the $20 per million output tokens that OpenRouter lists for Claude Opus 5.5. It is also about 2.1 times DeepSeek V4.1 Flash-1017’s $1.20 standard rate on the same marketplace.
The agent score is one row of 34 in Tencent’s table, and 20 of those rows are internal Tencent tests, by our count. On the other 14, Hy4’s score is above Claude Opus 5.5’s on two. Tencent writes that “a notable gap remains between it and top proprietary models”.
Business Pill · Open Weights
A short Business Engineer explainer on renting a model through an API versus running open weights yourself. It teaches the general idea only and does not discuss Tencent or Hy4.
The key insight: Tencent lists Hy4 at $2.501 per million output tokens, one-eighth of the $20 OpenRouter lists for Claude Opus 5.5, by our arithmetic, and its own table puts Hy4 at 75.2 on APEX-Agents v1.1 against 74.8 for Opus 5.5 in Tencent’s testing.
Update, 5 Nov, 04:09Z: Bloomberg reports Tencent’s claim that Hy4 joins China’s top tier of open-weight models. In Tencent’s own table, Hy4 tops all three Chinese rivals on 14 of 34 rows, 11 of them internal tests, by our count. Bloomberg News reported that Tencent “released a foundation model Wednesday that it said joins the top tier of Chinese open-weight AI systems for the first time”, and posted the story on X at 03:45 UTC on 5 November. Citing Tencent’s statement, Bloomberg wrote that Hy4 “has comparable performance to the latest offerings from Z.ai Co., Moonshot AI and DeepSeek in tests from coding to agents and browser use”. We read only the free opening of the article.
By our count of Tencent’s 34-row table, using Tencent’s own testing where a cell holds two figures, Hy4 scores above all three Chinese models listed, GLM-5.3, DeepSeek V4.1 Flash-1017 and Kimi K3.1 Pro, on 14 rows. Eleven of those 14 are internal Tencent benchmarks; the other three are ProgramBench, FrontierCode 2.0 and APEX-Agents v1.1. Against Kimi K3.1 Pro alone, Hy4 is higher on 16 rows, level on one and lower on 17. Against GLM-5.3 it is higher on all 29 rows where GLM-5.3 has a score. Every score is Tencent’s; the counts are our arithmetic.
Tencent’s page says “We plan to open-source Hy4’s model weights”. At 04:07 UTC on 5 November, a search of Tencent’s Hugging Face account for Hy4 returned only the Hy4-preview models, so the “open-weight” label rests on that stated plan.
Tencent’s Hunyuan account posted the launch on X at 06:43 UTC: “Hy4 is here, new with vision”. The release page says: “Today, weโre releasing Hy4, another major leap forward that places Hy among the strongest open-source models for the first time”.
The page says Hy4 “supports vision perception for the first time” and “also supports computer use for the first time”, across command-line and graphical interfaces. It lists Tencent Cloud and OpenRouter as API channels, and CodeBuddy, WorkBuddy, Yuanbao and ima as products that carry the model.
Tencent says Hy4 “delivers major performance gains over Hy4 preview at the same API price”. The weights are not out yet: “We plan to open-source Hy4โs model weights”, the page says, with a system card to follow at that point. It also plans to release Hy4 flash “in the near future”.
APEX-Agents v1.1 (pass@1) scores from Tencent’s own Hy4 benchmark table (Tencent’s testing where it gives two figures) against list price per million output tokens: Hy4 from Tencent’s release page, peers as listed on OpenRouter at 07:10 UTC on 4 November 2026 (DeepSeek at its standard rate). The 8x ratio is our arithmetic. Tencent’s scores have not been independently verified.
What Was Posted on X
๐Hy4 is here, new with vision. Firmly among the leading open models. More productive everywhere: finance, frontend, game, officework, researchโฆ Built responsibly, with cyber/biosecurity in mind. Try it out and let us know! ๐๏ผhttps://hy.tencent.ai/research/hy4?langVersion=en
Tencent’s Hunyuan account announced Hy4 at 06:43 UTC on 4 November, linking to the release page this piece is built on: “Hy4 is here, new with vision.” The post itself carries no price or benchmark figures; those are on the release page.
Tencents new Hy4 looks like a serious challenger to the strongest open-weight AI models.
In Tencentโs own tests, Hy4 scores 75.2 on APEX-Agents, ahead of DeepSeek V4.1 Flash-1017 at 72.1/72.9, but the two trade wins across other benchmarks.
Hy4 costs $0.834 per million input tokens and $2.501 per million output tokens via API, and Tencent says it plans to release the modelโs weights.
The X account @kimmonismus quoted Tencent’s post 11 minutes later with an image of Tencent’s benchmark table, citing Hy4’s 75.2 on APEX-Agents against DeepSeek V4.1 Flash-1017’s 72.1/72.9 and the $0.834 and $2.501 prices. Those figures match Tencent’s page. The post’s text does not discuss the Claude Opus 5.5 scores shown in the table, or the price comparison with Opus 5.5 that this piece is built on.
Tencent released Hy4, its new flagship model with performance among the top open models on benchmarks.
Hy4 is now available on Tencent Cloud, OpenRouter, CodeBuddy, WorkBuddy, and Hy AI Studio. Open weights are coming soon.
> “Tencent plans to open-source Hy4โs model weights and have partnered with the vLLM (Inferact Inc.) and SGLang communities to provide ecosystem support.
— ๐จ AI News | TestingCatalog (@testingcatalog) Nov 04 07:44
TestingCatalog reports that Tencent released Hy4, its new flagship model, and says its benchmark performance is among the top open models. The post lists availability on Tencent Cloud, OpenRouter, CodeBuddy, WorkBuddy and Hy AI Studio, says open weights are “coming soon”, and quotes Tencent’s plan to open-source the weights with vLLM (Inferact Inc.) and SGLang support and to publish a system card with safety evaluations when it does.
The Price, and the Launch Discount
Tencent’s price table lists $0.042 per million cached input tokens, $0.834 for input and $2.501 for output. OpenRouter’s public listing for Tencent’s Hy4 endpoint, read at 07:10 UTC, shows $0.417 for input and $1.2505 for output with a discount field of 0.5, which is half of Tencent’s list price, by our arithmetic. OpenRouter does not say how long the discount runs.
Against list prices on OpenRouter at the same time, Hy4’s $2.501 output price compares with $20 for Claude Opus 5.5, $10 each for GPT-6.1 Sol and Kimi K3.1 Pro, $4.40 for GLM-5.3 on Z.AI’s own endpoint and $1.20 for DeepSeek V4.1 Flash-1017 on DeepSeek’s endpoint. OpenRouter also lists lower DeepSeek rates of $0.60 at weekends and in set weekday hours.
By our arithmetic, Opus 5.5’s output price is 8.0 times Hy4’s, GPT-6.1 Sol’s and Kimi K3.1 Pro’s are 4.0 times, and GLM-5.3’s is 1.8 times. On input, Opus 5.5’s $4 is 4.8 times Hy4’s $0.834.
What Tencent’s Table Shows
Tencent’s notes say each model was run at its highest available reasoning setting, “with * indicating our own testing”. Where a cell holds two figures, the starred one is Tencent’s. On APEX-Agents v1.1, Tencent lists Claude Opus 5.5 at 73.5 and 74.8, DeepSeek V4.1 Flash-1017 at 72.1 and 72.9, Kimi K3.1 Pro at 68.2 in its testing, GLM-5.3 at 64.3 and GPT-6.1 Sol at 61.7 in its testing. Hy4 preview scored 53.9, so the new model is 21.3 points higher, by our arithmetic.
On the 14 rows that are not internal Tencent tests, Hy4 is above Opus 5.5 on APEX-Agents and DeepSWE v1.1, at 74.9 against 74.2 and 72.6, by our count. Opus 5.5 is ahead on the other 12, including ProgramBench, at 65.0 and 56.5 against 30.0, and Terminal-Bench 4.0, at 66.4 against 44.4.
Against DeepSeek V4.1 Flash-1017, Hy4 is ahead on 11 of the 14, behind on SWE-Marathon v1.1, at 43.8 against 46.3, and split on two where DeepSeek has two figures, by our count. Every score in the table comes from Tencent’s release.
Tencent’s Finance and Product Figures
Tencent says it evaluated Hy4 with 213 finance professionals on 579 real-world financial tasks, where it scored 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3. In internal evaluations, the page says, “Hy4 cut the hallucination rate from 15.76% (on Hy4 preview) to 5.45%”.
The page quotes Tencent product teams and outside firms. A statement attributed to Comein says “Hy4 is currently the only Chinese model to tie with Claude Opus 5.5 for second place overall” on its investment research agent tasks. Tencent’s game team for Honor of Kings is quoted saying configuration-task accuracy rose “from 91.74% to 97.28%”.
Safety Testing and the Weights
Tencent says it limited cybersecurity training data to defensive material and filtered out exploit development, reverse engineering and penetration testing. It reports a score of 30.4 out of 100 on its Hy-BioExperimentBench and writes: “We found no evidence that Hy4 meaningfully lowers the barrier to causing biological harm”.
Tencent says it has partnered with the vLLM and SGLang communities for the weights release. The page gives no date for it.
What the Post on X Said
Eleven minutes after Tencent’s post, the X account @kimmonismus quoted it with an image of Tencent’s table: “In Tencentโs own tests, Hy4 scores 75.2 on APEX-Agents, ahead of DeepSeek V4.1 Flash-1017 at 72.1/72.9, but the two trade wins across other benchmarks”. Its scores and prices match Tencent’s page. The post’s text did not discuss the Opus 5.5 scores shown in the table or the OpenRouter discount.
How Each Side Reads These Figures
For readers who see a price shift, the figures include an output price one-eighth of Opus 5.5’s, a 75.2 agent score above every other model in Tencent’s table, and a plan to publish the weights.
For readers who see caution, the figures include a table run by Tencent itself, 20 internal rows out of 34, Opus 5.5 ahead on 12 of the 14 other rows, an output price about 2.1 times DeepSeek’s standard rate, and weights with no release date. We cite no analysts’ views here.
As we read the table, the agent score is the exception rather than the pattern: on the 14 rows that are not internal Tencent tests, Opus 5.5 is ahead on 12, by our count.
As we read the prices, Hy4 sits between two groups on OpenRouter’s list: well below Opus 5.5, GPT-6.1 Sol and Kimi K3.1 Pro on output price, and about 2.1 times DeepSeek V4.1 Flash-1017’s standard rate.
Tencent, Hy4 release page, 4 November 2026
“Although Hy4 is one of the strongest open-weight models available today, a notable gap remains between it and top proprietary models”
Three Implications
FOR MODEL BUYERS Tencent lists $0.834 per million input tokens and $2.501 per million output tokens, and OpenRouter showed a 0.5 discount on Tencent’s endpoint at 07:10 UTC.
FOR WESTERN LABS Tencent’s table puts Hy4 at 75.2 on APEX-Agents v1.1 against 74.8 for Claude Opus 5.5, at one-eighth of Opus 5.5’s output price on OpenRouter, by our arithmetic.
FOR HOW THE SCORES ARE READ Every score is Tencent’s own, 20 of the 34 rows are internal Tencent tests, and the weights have no release date.
The framework states the pattern as: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize)”.
As we read it, Hy4 maps onto that split: Tencent plans to open the model weights, while WorkBuddy, Yuanbao, CodeBuddy and ima, the products that carry Hy4, remain Tencent’s own. The framework does not discuss Tencent or Hy4.
This piece is based on Tencent’s Hy4 release page dated 4 November 2026, including its benchmark appendix, which is an image we transcribed, and Tencent Hunyuan’s post on X. Every benchmark score is Tencent’s; we ran no tests and have not verified any score independently.
Peer prices are OpenRouter listings read at 07:10 UTC, not each lab’s own price page, and a model’s cost per task also depends on how many tokens it uses. The page as read does not give Hy4’s parameter count or a date for the weights. The one-eighth ratio and the row counts are our arithmetic. We did not contact Tencent.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Tencent released Hy4 on 4 November at $0.834 per million input tokens and $2.501 per million output tokens, and its own table puts the model at 75.2 on APEX-Agents v1.1, against 74.8 for Claude Opus 5.5 in Tencent’s testing. By our arithmetic, that output price is one-eighth of Opus 5.5’s on OpenRouter, while Opus 5.5 leads on 12 of the 14 non-internal rows in the same table.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Benchmark scores are Tencent’s own, from the benchmark appendix of its Hy4 release page dated 4 November 2026, which we transcribed from an image; we ran no tests. Hy4 prices are from Tencent’s page; peer prices are OpenRouter listings read at 07:10 UTC on 4 November 2026. Ratios and row counts are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
