Tencent Launches Hy4 flash at $0.07 per Million Tokens: Cheaper than Claude Haiku 5.5, but Behind DeepSeek in Its Own Tests
Tencent launched Hy4 flash on November 16, the budget version of its Hy4 model, priced at $0.07 per million input tokens and $0.281 per million output tokens, roughly twelve times less than Hy4. Tencent claims it matches Claude Haiku 5.5 at an 18 times lower cost per task. Yet its own charts show DeepSeek-V4.1-Flash ahead on quality, and the per-token price gap with Haiku 5.5 is far smaller than 18.

Twelve days after Hy4, Tencent is adding an entry-level version. Hy4 flash was announced on X on Monday, November 16, at 7:38 a.m. Paris time, as “the cheapest model for everyday productivity”. The headline claim fits in one sentence: the model would do as well as Anthropic’s Claude Haiku 5.5 at an 18 times lower cost per task. Tencent’s own published figures call for some nuance.
What is available
Tencent Cloud’s pricing page, updated on November 16, lists the model under the identifier hy4-flash on its TokenHub API platform. The prices match the announcement:
| Per million tokens | Hy4 | Hy4 flash | Gap |
|---|---|---|---|
| Input | $0.834 | $0.070 | ÷ 12 |
| Output | $2.501 | $0.281 | ÷ 9 |
| Cached input | $0.042 | $0.003 | ÷ 14 |
The same page grants a 50% discount on Hy4 flash until November 30 at noon Beijing time, and on Hy4 until December 31. During that period, the real gap between the two models therefore stays the same as at list price. Hy4 preview, the summer preview release, is marked “soon to be retired”.
Tencent recommends Hy4 flash for everyday tasks and keeps Hy4 for complex work such as long-running software engineering, according to Beijing News. The Beijing daily, a secondary source, also reports the technical specifications: 224 billion parameters, including 64 billion of “Engram” memory, and 12 billion active parameters, with a one-million-token window. Tencent reportedly plans to release the weights, with no date. These figures appear on Tencent’s technical page, which is protected against automated access and which we were unable to read: they remain to be confirmed.
“18 times cheaper than Haiku 5.5”: where the figure comes from
The figure comes from a blind test run by Tencent and published as an image in the announcement. Users rated the answers on a 4-point scale without knowing which model had produced them:
| Tencent’s blind test | Satisfaction (out of 4) | Median time | Cost per task |
|---|---|---|---|
| DeepSeek-V4.1-Flash-1017 | 3.09 | 10.3 min | 1.20 yuan |
| Claude Haiku 5.5 | 2.81 | 6.8 min | 4.73 yuan |
| Hy4 flash | 2.79 | 6.7 min | 0.26 yuan |
| GLM-5.3-Flash | 2.53 | 7.6 min | 0.81 yuan |
The “18 times” is the ratio between 4.73 and 0.26 yuan. Two points stand out in the table. First, the best-rated model is not Tencent’s but DeepSeek-V4.1-Flash, clearly ahead (3.09 versus 2.79), even though it is slower and costs about four times more per task. Second, the tie with Haiku 5.5 comes down to two hundredths of a point, in a test designed by Tencent.
Per-token pricing alone cannot explain a factor of 18 either. According to Anthropic’s price list, Claude Haiku 5.5 costs $0.10 for input and $0.50 for output per million tokens for prompts up to 100,000 tokens. At that level, Hy4 flash is only 30% cheaper for input and 44% cheaper for output. Beyond 100,000 tokens of context, common for an agent, Haiku 5.5 rises to $0.50 and $2.50: the gap then grows to 7 times for input and nearly 9 times for output. The rest of the gap comes down to the number of tokens consumed per task, which Tencent does not publish.
The limits Tencent itself shows
On its internal WorkIndex, Tencent gives Hy4 flash 42.1, versus 51.4 for Hy4, at a cost per task of about 0.2 yuan versus 5.7: 24.5 times cheaper according to Tencent, but clearly weaker. On the public benchmarks chart, Hy4 flash also sits below DeepSeek-V4.1-Flash and Qwen3.8-Flash-Next, while being the cheapest per task. According to TokenPost, a secondary source, it scores 19.2 on Terminal-Bench 4.0, versus 44.4 for Hy4: a model for simple tasks, not for demanding coding agents.
On TokenHub, DeepSeek-V4.1-Flash is billed at $0.15 for input and $0.60 for output during off-peak hours, and double that during weekday peak hours, Beijing time. Hy4 flash therefore costs roughly two to four times less per token than its best-rated competitor on the same platform.
Our take
Hy4 flash is not trying to be the best small model, and Tencent’s own charts show it: DeepSeek-V4.1-Flash stays ahead on quality. Its argument is cost per task, and that rests partly on price and partly on lower token consumption, which Tencent claims without detailing it. The “18 times cheaper than Haiku 5.5” is therefore a use-case figure, not a price: per token, the gap with Haiku 5.5 ranges from 30% to about 9 times depending on context length. For a team handling large volumes of simple tasks (summaries, extraction, documents), the model is worth a try, especially since the discount runs until November 30. On quality, independent measurements are needed, and for local hosting, the release of the weights, which Tencent itself has not yet announced.
Sources
- Tencent Hy (X) — Meet Hy4 flash, the cheapest model for everyday productivity Primary source
- Tencent Hy (X) — WorkIndex and cost per task (announcement thread) Primary source
- Tencent Hy (X) — WorkIndex and public benchmarks, cost per task (announcement thread) Primary source
- Tencent Cloud — TokenHub model pricing (USD) Primary source
- Anthropic — Pricing (Claude documentation) Primary source
- Beijing News (新京报) — 腾讯发布Hy4 flash:输入0.5元/百万Token,计划开源 Secondary source
- TokenPost — Tencent's Hy4 Flash Cuts Automated-Task Cost to 0.038 Yuan Secondary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.


