770B / 49B ACTIVE1M CONTEXTPREVIEW, NOT GA95% CACHE DISCOUNTOPEN WEIGHTS
Hunyuan Hy4 Preview API Pricing
Hy4 Preview is the largest model Tencent has ever put on a price list - a 770B total / 49B active MoE with a 1M-token context, open-sourced August 28, 2026. Tencent Cloud International lists it at $0.834/M input, $2.501/M output, and $0.042/M cache hit. Those are the vendor's own USD figures, not a conversion of the CNY card - Tencent publishes the two lists independently.
Input - per 1M tokens
$0.834/M
6.3x the Hy3 rate vs Hy3
Output - per 1M tokens
$2.50/M
4.7x the Hy3 rate vs Hy3
Cache hit - 95% off
$0.042/M
Only 1.3x the Hy3 cache rate -95%
Effective - agentic blend
$0.37/M
92/8 split - 82% cache
§ 01 / TERMINAL
Run the numbers.
Live calculator pre-loaded with Hy4 Preview's Tencent Cloud International rate. The cache slider matters more here than on any other Hunyuan row - the cache hit is 95% below fresh input, so the blended cost swings by a factor of four across the three presets.
$ /mo
Workload split
Prompt cache hit rate
Tokens you can process
—
Words equivalent (English)
—
Effective rate
—
§ 02 / SCENARIOS
Real-world presets.
CODING AGENT
Codebase-scale migration
$0.047/task
LONG DOC ANALYSIS
Reading 100-page contracts
$0.053/doc
RAG - SUPPORT
Ticket triage
$0.004/ticket
ASSISTANT
Research planning turn
$0.008/turn
§ 03 / TOKENIZER
Paste text. See tokens. See cost.
Calibrated · measured on the vendor's tokenizer · 2026-06-10 Auto-counts as you type
Counts use a chars-per-token calibration measured on the vendor's own published tokenizer (tencent/Hunyuan-A13B-Instruct, 2026-06-10). English prose is typically within a few percent; code and non-Latin scripts tokenize heavier. For billing-exact counts use the vendor's count-tokens API.
Characters 704
Words 118
Tokens (estimated) 135 tokens
Cost as input · uncached $0.00011 USD
Cost as output · uncached $0.00034 USD
Cost as cached input $0.00001 USD
| Model | Input /M | Output /M | Effective blended | Context | Best for |
|---|---|---|---|---|---|
| Hunyuan Hy4 Preview Current | $0.834 cache $0.042 | $2.50 | $0.37 agentic 92/8 | 1M | Tencent's largest model - long-context agentic work |
| Hunyuan Hy3 | $0.132 cache $0.033 | $0.528 | $0.089 the GA flagship it sits above | 256K | Production traffic at a sixth of the input rate |
| Hunyuan A13B | $0.0744 | $0.298 | $0.0923 no cache rate published | 224K | Cheapest Tencent row still on the price list |
| DeepSeek V4 Pro | $1.32 cache $0.044 | $3.96 | $0.569 peak-hour rate | 1M | Chinese frontier reasoning at 1M context |
| GLM-5.3 | $1.40 cache $0.26 | $4.40 | $0.78 shallower cache discount | 1M | Coding-first frontier alternative |
| Kimi K3 | $3.00 cache $0.30 | $15.00 | $1.92 output-heavy pricing | 1M | Long-horizon agentic reasoning |
Frequently asked.
What the preview label actually costs you, why the rate went up rather than down, and where the USD figures come from.
Q · 01 What does Hy4 Preview cost? +
Tencent Cloud International lists
$0.834/M input, $2.501/M output, and $0.042/M on a cache hit, for the Singapore region. Tencent prints the same three figures in its own launch announcement, so the price list and the press release agree - which is not always true of a launch-week model.Q · 02 Why is the new model more expensive than Hy3? +
Because it is a different weight class, not a refresh. Hy4 Preview is
770B total parameters against Hy3's 295B, and the price moved with it: input is 6.3x the Hy3 rate and output 4.7x. A generational rise is unusual - most vendors cut on the successor - so read this as Tencent adding a tier above Hy3 rather than replacing it. Hy3 is still on the same price list at the same rate.Q · 03 Does the cache discount change the picture? +
Substantially. The cache-hit rate is
$0.042/M, only 1.3x Hy3's $0.033/M, while fresh input is 6.3x. So on cache-heavy agentic traffic the two models are 4.2x apart on the blended rate rather than 6.3x. The 95% gap between fresh and cached input is the widest on any Hunyuan row - if your prompts do not reuse a prefix, this model gets expensive fast.Q · 04 How does it compare with other Chinese frontier models? +
On this site's agentic blend it is the cheaper side of every 1M-context frontier row we price from a Chinese lab:
$0.370/M against DeepSeek V4 Pro at $0.569 (peak-hour), GLM-5.3 at $0.780, and Kimi K3 at $1.923. Those are blended figures under one shared assumption, not benchmark scores - the shelf below shows the raw rates.Q · 05 What does the preview label mean for production use? +
Tencent sells it as a preview row on TokenHub, so we carry it with
preview status rather than GA. Practically: the rate is real and billable today, but Tencent has a recent habit of dating preview rows out - the previous Hy3 preview row was scheduled offline for August 31, 2026 while GA Hy3 continued unaffected. Plan a fallback to Hy3.Q · 06 Is it really open weight, and under what licence? +
Tencent's own announcement is headlined "Releases and Open-Sources", and ships a checkpoint, an FP8 quantised variant and serving recipes. The announcement does not name a licence, and we have not been able to open the weight repository to confirm one - so this page says open weights and stops there rather than repeating a licence figure from second-hand coverage.
Q · 07 Where do the USD figures come from? +
Tencent Cloud International's own USD price list, not a conversion. The CN-region card prices this model in yuan (
Y6 / Y18 / Y0.3), and converting it would give $0.845 / $2.535 - close, but wrong. Tencent applies no single exchange rate across the two lists: Hy4's USD implies about Y7.19/USD while Hy3's implies Y7.58/USD. Where a vendor publishes USD, we read it.