Tencent has released Hy4 flash, a cheaper version of its Hy4 model, and says it cost 1/18th as much per task as Anthropic’s Claude Haiku 5.5 in its own blind test, by our arithmetic from Tencent’s figures: 0.26 yuan per task against 4.73 yuan. In the post dated 16 November, Tencent puts Hy4 flash at 2.79 for user satisfaction on a four-point scale, against 2.81 for Haiku 5.5.
On list prices the gap is far smaller. Tencent lists Hy4 flash at $0.281 per million output tokens, and Anthropic lists Haiku 5.5 at $0.50 for prompts up to 100,000 tokens, about 1.8 times as much, by our arithmetic.
Business Pill · Tokens Explained
An evergreen Business Engineer explainer on how AI models are metered and priced in tokens. It teaches the general idea only and does not discuss Tencent or Hy4 flash.
The key insight: As we read it, Tencent’s own figures put most of the saving outside the rate card: Haiku 5.5’s output list price is about 1.8 times Hy4 flash’s for prompts up to 100,000 tokens, but its cost per task in Tencent’s blind test was 18.2 times, by our arithmetic. The satisfaction scores, 2.79 and 2.81, come from Tencent’s test of its own users.
What Tencent Released
Tencent’s post calls Hy4 flash “our most cost-effective Hy model to date”, with 224 billion total parameters and 12 billion active per token. It says the model is natively multimodal with a one-million-token context window. Tencent’s research page dates the flagship Hy4 to 3 November, and we covered its pricing against Claude Opus on 4 November.
The post lists Hy4 flash at $0.070 per million input tokens, $0.281 per million output tokens and $0.003 for cached input, against $0.834, $2.501 and $0.042 for Hy4. Tencent Cloud’s TokenHub page shows the same list prices at a limited-time 50% discount, and a context window of 1024K tokens.
Tencent says the model is available in its WorkBuddy, Yuanbao, WeCom and QQ apps and through TokenHub, and that it “will launch on OpenRouter soon”. The post says WorkBuddy users can use it for free through 30 November, and that Tencent plans to release Hy4 flash as an open-weight model.

Tencent’s Blind Test
The post describes “a blind evaluation with 85 internal WorkBuddy users across 259 tasks spanning 27 categories”. Its results chart compares four models on satisfaction, median end-to-end time and cost per task, and its note says the external models used their high tier.
In that chart, DeepSeek-V4.1-Flash-1017 scored highest on satisfaction, at 3.09, at a cost of 1.20 yuan per task. Haiku 5.5 scored 2.81 at 4.73 yuan, Hy4 flash 2.79 at 0.26 yuan, and GLM-5.3-Flash 2.53 at 0.81 yuan. Median completion time was 6.7 minutes for Hy4 flash, 6.8 minutes for Haiku 5.5 and 10.3 minutes for the DeepSeek model.
Tencent sums up the result as user satisfaction “close to Claude Haiku 5.5”. The gap is 0.02 points on the four-point scale, and DeepSeek’s model scored 0.28 points above Haiku 5.5, by our arithmetic. As we reported on 16 November, DeepSeek began rationing its free chat service, capping input at 40,000 characters.

Price per Token vs Cost per Task
Anthropic’s pricing page lists Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, rising to $0.50 and $2.50 for longer prompts. Against Hy4 flash’s single list price, that is about 1.4 and 1.8 times for shorter prompts, and about 7.1 and 8.9 times for longer ones, by our arithmetic.
The 18.2-times gap in Tencent’s test is about 10 times the 1.8-times gap in output list price, by our arithmetic. Cost per task also depends on how many tokens a model uses, which price tier its prompts fall into, and caching. Tencent says reinforcement learning “cuts average end-to-end output tokens by 50%”.
Tencent also says that on WorkIndex, its internal benchmark, Hy4 flash’s average cost per task is 1/20 of Hy4’s and 1/8 of DeepSeek-V4.1-Flash-1017’s at that model’s Max tier and peak-hour pricing. Alibaba made a cost-per-task argument too, when it launched Qwen4-Max at $4.50 per million output tokens, citing a fall from $4.43 to $1.67 on its own Work360 test.
The Structural Read
As we read Tencent’s chart, the cheapest model was not the best rated: DeepSeek-V4.1-Flash-1017 scored 3.09, the highest of the four, at 1.20 yuan per task, about 4.6 times Hy4 flash’s cost and about a quarter of Haiku 5.5’s, by our arithmetic.
As we read the post, Tencent measures the model by cost per task, not price per token, as Alibaba did for Qwen4-Max. On that measure the number of tokens a model spends on a task counts alongside the rate card, and Tencent says reinforcement learning cut average end-to-end output tokens by 50%.
Tencent Hy, Introducing Hy4 flash, 16 November 2026
“We recommend Hy4 flash for most everyday productivity tasks—such as deep search, document processing, slide creation, simple frontend demos, and computer use—while continuing to recommend Hy4 for complex workloads like long-horizon software engineering for its higher intelligence ceiling.”
Three Implications
FOR DEVELOPERS Tencent lists Hy4 flash at $0.070 per million input tokens and $0.281 per million output tokens, and says TokenHub API prices are 50% off and WorkBuddy access is free through 30 November.
FOR ANTHROPIC Its pricing page lists Haiku 5.5 at $0.10 and $0.50 per million tokens for prompts up to 100,000 tokens; Tencent’s note says the external models in its test used their high tier.
FOR OPEN MODELS Tencent says it plans to release Hy4 flash as an open-weight model and that the model will launch on OpenRouter soon.
The Business Engineer Lens
This story maps onto the Business Engineer framework The State of The Inference Economy.
The analysis puts it this way: “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, Tencent’s cost-per-task chart is a measure of that marginal cost: 0.26 yuan per task for Hy4 flash against 4.73 yuan for Haiku 5.5 in Tencent’s own test. The analysis was published in March 2026 and does not discuss Tencent or Hy4 flash.
What Is Not Established
Every test result here is Tencent’s own: its own users, its own task set and its own cost accounting, reported in yuan. We have not run the test. We could not split the 18-times gap into price and token use from the figures Tencent published.
The list-price comparison uses standard prices before batch discounts and before Tencent’s temporary 50% TokenHub discount. We did not contact Tencent or Anthropic for this piece.
The Bottom Line
In Tencent’s own blind test, Hy4 flash scored 2.79 against Haiku 5.5’s 2.81 and cost 0.26 yuan per task against 4.73 yuan, about 1/18th, by our arithmetic. On the two companies’ list prices for prompts up to 100,000 tokens, Haiku 5.5’s output costs about 1.8 times as much.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece rests on Tencent’s post announcing Hy4 flash, which we read in full, including its benchmark and blind-test figures, and on Tencent Cloud’s TokenHub page for the model. Every test result is Tencent’s own. Haiku 5.5 prices come from Anthropic’s own pricing page, read on 16 November 2026. TechNode’s report was used for context. The ratios and point gaps are our arithmetic. We did not contact the companies named.
Sources: Tencent Hy, Introducing Hy4 flash, 16 November 2026 · Tencent Cloud TokenHub, Hy4 flash (read 16 November 2026) · Anthropic, Pricing (read 16 November 2026) · TechNode, Tencent launches Hy4 flash, 16 November 2026









