Early-bird pricing (ends Dec 25), locked in while you stay subscribed: Go $10/mo (normally $20), Plus with Opus 5.5 just $30/mo (normally $50)Subscribe →

Hunyuan Hy4 Flash API Pricing: $0.07 Input, 224B, 1M Context

Tencent's Hy4 Flash (224B total / 12B active, 1M context) is live as of November 16, 2026. Standard rate is CNY 0.50 input / CNY 2.00 output per million tokens ($0.070 / $0.281); the current 50% TokenHub launch rate is CNY 0.25 / CNY 1.00 through November 30 at noon Beijing time.

Flash is not a smaller Hy4 for the same jobs. Billing follows the ID you send: hy4-flash is the cheap tier, hy4 the flagship, hy4-preview the August checkpoint.

Hy4 Flash is live — what shipped on November 16

The November 4 Hy4 announcement left Flash unpriced; the November 16 release post and same-day TokenHub listing closed that gap (ITHome, Sina Finance, LLMLearner specs).

  • Tencent Cloud TokenHub, model ID hy4-flash — live.
  • WorkBuddy, Yuanbao and Tencent apps — with WorkBuddy free through November 30.
  • OpenRouter — planned, not live. Only tencent/hy4 is listed as of November 16; use TokenHub's Singapore endpoint until OpenRouter adds Flash.

Specs that explain why Flash costs one-tenth of Hy4

Hy4 Flash totals 224B parameters including 64B Engram parameters, with 12B active per token, in a mixture-of-experts layout (LLMLearner, Sina). Context window is 1M tokens (1,048,576), input is text plus image, output is text.

What matters for the bill: KV cache is 1/6 of Hy4 and 1/24 of Hy3, with megakernel, 4-bit quantization and dflash acceleration for decoding above 150 tokens/s. Multimodal image handling, chart parsing and visual math run on the same endpoint.

Weights are announced, not shipped: unlike Apache 2.0 Hy4-preview, Flash has no public checkpoint, date or license as of November 16.

Hy4 Flash API pricing in one table

TokenHub publishes separate mainland (CNY) and international (USD) catalogs — regional schedules, not conversions. Flash promo runs to November 30 at 12:00 Beijing time (LLMLearner); flagship promo runs to December 31 at 23:59 Beijing time (CostMyTokens verification, ChinaModelAPI).

TierInput / 1MOutput / 1MCache read / 1MValid
Flash list (CNY)¥0.50¥2.00¥0.02standard
Flash promo (CNY)¥0.25¥1.00¥0.01to Nov 30, 12:00 Beijing
Flash list (USD)$0.070$0.281$0.003standard
Hy4 list (CNY)¥6.00¥18.00¥0.30standard
Hy4 promo (CNY)¥3.00¥9.00¥0.15to Dec 31, 23:59 Beijing
Hy4 list (USD)$0.834$2.501$0.042standard
Hy4 promo (USD)$0.417$1.2505$0.021to Dec 31
Hy4 Flash vs Hy4 TokenHub pricing in CNY per million tokens

Worked example at Flash promo: 100K fresh input + 500K cached input + 5K output = 0.1 x ¥0.25 + 0.5 x ¥0.01 + 0.005 x ¥1.00 = ¥0.035.

Per-task cost is the real gap. Tencent reports roughly 1/20 of Hy4, 1/13 of Hy4 preview and 1/8 of DeepSeek-V4.1-Flash (max) (Sina). Community math on the launch thread, derived from Tencent's blind test of 85 internal users over 259 tasks (LLMLearner), put it at ¥0.26 per task versus ¥4.73 for Claude Haiku 5.5 (high) and ¥1.20 for DeepSeek-V4.1-Flash, satisfaction 2.79 vs 2.81 (@alphaw0ng breakdown). No independent table exists yet — treat it as a vendor test with community arithmetic.

Cache decides agent bills: cache-read is 4–5% of fresh input (¥0.02 vs ¥0.50; $0.003 vs $0.070).

"Is it cutting-edge model? No, not at all. But this is the key to makes workbuddy the Doubao of the working agent. Just 200B-parameter means they can flood the mass market with very low-cost tokens enough to handle relatively simple tasks." — @LeroyLi311063, launch-day reaction

Flash vs Hy4 vs Hy4 Preview: pick by workload

Hy4 FlashHy4 (GA, Nov 4)Hy4 Preview (Aug 28)
Params224B total / 12B activeundisclosed (preview ref: 770B / 49B)770B / 49B
Context1M1M — 960K max input, 64K max output (setup guide)1M
Modalitiestext + image in, text outtext + image in, text out, computer usetext, early checkpoint
TokenHub IDhy4-flashhy4hy4-preview
List input/output¥0.5 / ¥2.0¥6 / ¥18 ($0.834 / $2.501)¥6 / ¥18
Best forsearch, docs, slides, simple frontend, everyday computer uselong-horizon software, complex agents, finance researchtesting open weights (Apache 2.0)

Flagship context matters because Flash sits below it. In Tencent's Hy4 benchmark appendix, Hy4 trails top proprietary runs on Terminal-Bench 4.0 (44.4 vs ~66 for Opus 5.5-class) and ProgramBench (30.0 vs 56.5+). For SWE-bench-style or multi-hour terminal agents, the ¥0.25 promo input price does not make Flash the right model. Flagship rates are tracked in our Hy4 Preview API pricing guide and cheapest LLM API.

Call hy4-flash in 10 minutes

TokenHub is OpenAI-compatible with /v1/chat/completions, /v1/responses and Anthropic /v1/messages (endpoint guide). Use a TokenHub key, not a Yuanbao or WorkBuddy session token. Pick Guangzhou https://tokenhub.tencentmaas.com/v1 or Singapore https://tokenhub-intl.tencentmaas.com/v1, enable the model service in the console — a disabled service returns 402 per the same guide — then send hy4-flash.

bash
curl https://tokenhub-intl.tencentmaas.com/v1/chat/completions \
  -H "Authorization: Bearer $TOKENHUB_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "hy4-flash",
    "messages": [{"role": "user", "content": "Summarize this 80-page PDF in 10 bullets"}]
  }'
python
from openai import OpenAI
client = OpenAI(api_key="YOUR_TOKENHUB_API_KEY",
                base_url="https://tokenhub-intl.tencentmaas.com/v1")
resp = client.chat.completions.create(
    model="hy4-flash",
    messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)

Flash max output is undisclosed; Hy4 documents 64K — do not assume Flash inherits it. Check the console model page before planning 100K+ generations.

The routing rule: Flash by default, Hy4 for the hard 5%

If your workload is…UseWhy
Docs, slides, spreadsheets, meeting prep, deep searchhy4-flashNear-Haiku satisfaction at ~¥0.26/task in Tencent's test
High-volume agents with stable long prefixeshy4-flash4–5% cache ratio compounds fast
Simple frontend, everyday computer usehy4-flashPositioned here by Tencent; >150 tps helps interactivity
Long-horizon coding, terminal agents, program synthesishy4Flash is explicitly not positioned here
Self-hosting todayhy4-preview weightsOnly preview has downloadable Apache 2.0 weights