Skip to main content
TokenCost logoTokenCost
Model ReleaseNovember 22, 2026·7 min read

Tencent Hy4 Flash lists at $0.07 in and $0.281 out, and bills half that until noon Beijing time on November 30. At list it runs a cached support bot for 57% of GPT-6 Luna's bill and 48% of DeepSeek V4.1 Flash's off-peak one.

The cheapest card in Tencent's lineup, with a discount that ends a month before the one on its big sibling.

Branching lightning bolt glowing purple inside dark storm clouds at night

Photo by Breno Machado on Unsplash

Tencent released Hy4 Flash on November 16, twelve days after Hy4 itself. The TokenHub price page lists it at ¥0.5 per million input tokens, ¥0.02 on a cache hit and ¥2 output. In dollars, on TokenHub and on OpenRouter, that's $0.07 / $0.003 / $0.281.

Neither of those is what you pay this week. A notice on the same page bills Flash at 50% of list until 12:00 Beijing time on November 30, which is 04:00 UTC. OpenRouter already charges $0.035 / $0.0015 / $0.1405 and returns a discount: 0.5 field from its endpoints API.

The weights are public too. The model card describes a mixture-of-experts of about 220B parameters, 64B of them Engram memory, with 12B active per token, a 1,048,576-token context and a custom Tencent Hy License.

The card, in both currencies

Per 1M tokensYuan listDollar listBilled until Nov 30
Input¥0.5$0.07$0.035
Cache hit¥0.02$0.003$0.0015
Output¥2$0.281$0.1405
Cache write---

Tencent TokenHub pricing (page updated November 16) and the OpenRouter endpoints API, read November 22, 2026. No batch row for Flash, no length tiers.

A cache hit is 4% of the input price. That's a steeper discount than the 10% OpenAI and Anthropic charge on Luna and Haiku 5.5, and it's the line that matters for agents that resend the same system prompt every turn.

The dollar card isn't a live conversion. It works out to about ¥7.13 per dollar, close to the ¥7.2 Tencent used for Hy4. At Sunday's spot rate of 6.69, the yuan card is $0.0747 in and $0.299 out, so mainland customers pay roughly 6% to 7% more than dollar customers for the same tokens.

Two discounts, two end dates

Hy4 has its own half-price promotion, and it runs to December 31. Flash's ends on November 30. For one month the cheap model is at full price and the expensive one isn't, which changes Tencent's favorite comparison.

Tencent says a task on its WorkIndex benchmark costs 1/20 of what it costs on Hy4. Per token, Flash is 1/12 of Hy4 on input, 1/14 on cache hits and 1/9 on output. Getting from there to 1/20 means Flash uses roughly half Hy4's tokens per task, which fits the card's claim that reinforcement learning cut end-to-end output by 50%.

Run the same task in December and the ratio is closer to 1/10. Still a big gap. Just not the one in the launch post, and anyone benchmarking the two during December should check which price each one was billed at.

A support bot's monthly bill

Take a support agent handling 100,000 conversations a month. Each call sends a 16,000-token system prompt and knowledge base that sits in cache, 4,000 fresh tokens of conversation, and gets 1,500 tokens back. Same tokens for every model.

Hy4 Flash (until Nov 30)
$37.47
Hy4 Flash (list)
$74.95
GPT-6 Luna / Claude Haiku 5.5
$131.00
Qwen3.8 Flash
$154.50
DeepSeek V4.1 Flash (off-peak)
$154.80
DeepSeek V4.1 Flash (peak)
$309.60
Hy4 (half price to Dec 31)
$387.98
Hy4 (list)
$775.95

Our arithmetic from the Tencent, OpenAI, Anthropic, Alibaba and DeepSeek price pages. Cache reads only, no cache write fees counted.

At list, Flash runs that bot for $74.95. GPT-6 Luna and Claude Haiku 5.5 share a card and both come to $131, so Flash is 57% of them. DeepSeek V4.1 Flash off-peak and Qwen3.8 Flash land within 30 cents of each other at about $155, so Flash is 48% of either. Until November 30 it's $37.47.

Flash isn't the floor of the market, though. Qwen3.7 Flash lists at $0.03 / $0.13 and Ling-3.0-flash at $0.06 / $0.18. What Flash adds at its price is the 1M context, image input and open weights, and Tencent's blind test, where users rated it 2.79 out of 4 against Haiku 5.5's 2.81, at a reported ¥0.26 (about $0.04) a task.

What "1/8 of DeepSeek" is measured against

The model card's other claim is that Flash costs 1/8 as much per task as DeepSeek-V4.1-Flash-1017. The fine print says that's DeepSeek at its max reasoning tier and its peak-hour price. DeepSeek doubles every line between 01:00-04:00 and 06:00-10:00 UTC on weekdays, which is about a fifth of the week.

Against peak, Flash's list price is 23% of DeepSeek's on input and output. At this week's half price it's 11.7%, close to 1/8.5 with no token savings at all. Against off-peak, which is most of the week, Flash at list is 47% of DeepSeek per token, and cache hits cost the same $0.003.

So Flash is cheaper than DeepSeek's budget model either way. It's about twice as cheap per token for most traffic, not eight times. Any extra comes from how many tokens each spends, and no independent benchmark has measured that yet. Artificial Analysis lists Hy4 at an index of 35 but has no Flash page as of today.

Long prompts are where the flat card pays

Flash charges the same rate from the first token to the millionth. Luna reprices the whole request past 272K input tokens and Haiku 5.5 past 100K, so a large prompt sees the biggest gaps.

Send 800,000 uncached tokens of logs and ask for 4,000 back. Flash at list charges $0.057. DeepSeek V4.1 Flash off-peak charges $0.122. Luna, on its long-context rate of $0.20 / $0.75, charges $0.163, and Haiku 5.5 at $0.50 / $2.50 charges $0.41, about seven times Flash.

Our read

If you run cached, high-volume chat or classification and can send traffic to a Chinese provider, Flash is worth a test this week while it's at $0.035 / $0.1405. Plan the budget at $0.07 / $0.281. That's the price from November 30.

If you already use Hy4 and you're tempted to move everything to Flash, note December. Hy4 stays at half price until New Year, so the savings from switching shrink for a month before they return.

And treat the 1/8 and 1/20 ratios as Tencent's numbers on Tencent's benchmark until someone else measures them. The per-token card is the only part you can verify today, and on that Flash is about half of DeepSeek and Qwen, and a bit over half of Luna and Haiku.

We covered Hy4 and Hy4 preview, and the budget cards it competes with in DeepSeek V4.1 Flash, GPT-6 Luna and Claude Haiku 5.5. To run your own token mix, use the calculator.

Sources