Independent · Source-cited · No sponsored rankingsSources linked in every article
AIToolRanked
AI News · 5 min read

Hy4 flash is $0.07 in and $0.281 out

On November 16, 2026, Tencent listed Hy4 flash at $0.070 input and $0.281 output per million tokens. The savings figures are Tencent's own test.

RA
· Founder, AIToolRanked
A dark office at night with a long translucent ribbon across a desk ending in a single point of light
On this page

What Tencent posted on November 16

On November 16, 2026, Tencent Hy posted Introducing Hy4 flash ↗. The post says the model has 224 billion total parameters, including 64 billion Engram parameters, and 12 billion active per token. It is natively multimodal, with a 1 million token context window, at a lower API price than Hy4.

The picture above is ours. We have not called the model.

Tencent's Hy API guide ↗ lists hy4-flash at context 1M, maximum input 960k, and maximum output 64k, on the OpenAI Chat Completions API, OpenAI Responses, and the Anthropic Messages API. The example host is tokenhub-intl.tencentcloudmaas.com. The model list ↗ adds reasoning, structured output, function calling, caching, and image understanding. The guide says reasoning depth defaults to high.

The post puts it on TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with OpenRouter "soon," and says an open-weight release is planned. No date. No file.

What the price pages charge

The blog's table is per million tokens.

ModelCached inputInputOutput
Hy4 flash$0.003$0.070$0.281
Hy4$0.042$0.834$2.501

International model pricing ↗, updated November 16, 2026 at 12:15, prints $0.07, $0.281, and $0.003 for hy4-flash. China TokenHub pricing ↗, updated the same day at 12:15, prints Guangzhou at ¥0.5, ¥2, and ¥0.02.

Until 12:00:00 on November 30, 2026 (Beijing time), Hy4 flash is billed at 50% of list, only for usage inside that window. Hy4 stays at 50% of list until 23:59:59 on December 31, 2026 (Beijing time). While both discounts run, the list-price ratio is the bill ratio.

The blog says WorkBuddy is free through November 30. WorkBuddy's own post describes a quota. WorkBuddy ↗: "Free to try for a limited time" and "Daily free quota applies; then credits kick in."

On the international page, DeepSeek-V4.1-Flash-1017 (Vendor Direct, id deepseek/deepseek-flash) is $0.15 in and $0.60 out off-peak, and $0.30 and $1.20 at peak, with cache $0.003 and $0.006. Hy4 flash's list is lower than that off-peak rate. Our Ultrafast note is a different endpoint, at $2.40 and $9.60.

Gemini 4 Hydrogen is $1.50 and $7.50 on Google's November 12 post. Qwen4-Max is $1.50 and $4.50 on Alibaba's Singapore list as of November 16. Hy4 flash's list output is under both. Hydrogen's post allows up to 1M output. The Hy guide stops hy4-flash output at 64k and input at 960k.

What Tencent's tests claim

Every figure here is Tencent's. We did not run the tests.

The post says decoding passes 150 tokens per second. On WorkIndex, its internal office-task set, Tencent says average cost per task is 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 at Max tier, peak-hour pricing. A later paragraph says task time falls 70% at roughly 1/15 the cost of Hy4. Both ratios are in the post. They do not match. TechNode ↗'s "95%" headline is the one-twentieth line, not a second test.

The blind test is 85 internal WorkBuddy users, 259 tasks, 27 categories. The chart's 4-point satisfaction, median time, and cost: DeepSeek-V4.1-Flash-1017 3.09, 10.3 min, ¥1.20; Hy4 flash 2.79, 6.7 min, ¥0.26; Claude Haiku 5.5 2.81, 6.8 min, ¥4.73; GLM-5.3-Flash 2.53, 7.6 min, ¥0.81. The chart note says external models used the high tier. DeepSeek leads satisfaction. Hy4 flash leads cost. The time gap versus Haiku is 6.7 minutes to 6.8.

Tencent says to use Hy4 flash for deep search, documents, slides, simple frontend demos, and computer use, and to keep full Hy4 for long-horizon software engineering.

What people are saying

  • "HY4 family is working with DS Flash, instead of trying to compete with - This is a great strategy" — Tiezhen WANG ↗ (@Xianbao_QIAN), November 16. He also writes that it "will be open source soon." The blog says "plan" and names no date.

  • "but honestly, I still prefer Haiku 5.5's generation." — Sahil Panhotra ↗ (@SahilPanhotra), November 16, one peacock SVG prompt. His clock: Haiku 5.5 "6 minutes" and "$0.17 cost ($0.24 without cache)"; Hy4 flash "5 minutes 27 seconds," "110–120 tokens/sec," "$0.0244 cost." He adds "around 7x cheaper" and "currently FREE on @opencode." The free OpenCode line is not on Tencent's price page.

  • "personally i think hy4 ~ ds4.1 > hy4 flash ~ hy4 preview, but cost wise hy4 flash < ds4.1 << hy4 preview < hy4" — Shunyu Yao ↗ (@ShunyuYao12), November 16. The profile says "Language agents." It does not name an employer.

  • "Your 128GB Mac was never too small for a frontier MoE. The MoEs were too fat." — Sebastian Christiansen ↗ (@CerebralCoding_), November 16. He adds "about 80GB of real weights at 4-bit" and "Weights aren't public yet." The 80GB figure is his estimate.

Reddit was not reachable for this story.

Our take

Call it when the token bill is the constraint and a Tencent Cloud account is acceptable. List price at $0.070 and $0.281 sits under DeepSeek's vendor-direct off-peak rate on the same page, and under the Hydrogen and Qwen4-Max output rates above. The half-off until noon Beijing on November 30, and the WorkBuddy quota in the vendor's own post, are the window to score a real task.

Do not treat the chart as a quality win, or the weights as a download. DeepSeek leads Tencent's own satisfaction column. The 1/20 and "roughly 1/15" lines disagree, and both are self-reported. Tencent still sends long-horizon coding to full Hy4. Open weights are a plan. Until a file exists, start from the open-source LLM comparison. For a coding agent under Claude's list rate, use the cheaper Claude Code alternatives, then measure hy4-flash on your own traces. Output stops at 64k. Input stops at 960k.

Who should try it: teams on high-volume search, documents, slides, and computer-use steps, calling TokenHub, who will score their own tasks before November 30. Who should skip it: anyone who needs weights this week, a 1M output cap, or a coding lead Tencent does not claim.

More from the news desk · About Rai Ansar