LONG AI
Home › Blog › Tencent Hy4 Preview Retires Dec 7

Tencent Hy4 Preview Retires Dec 7

Category:资讯 · 6 min read · 2026-11-10

On Nov 9, Tencent Cloud posted a retirement notice across three platforms: Hy4 preview (model value hy4-preview) stops service at 00:00:00 Beijing time on 2026-12-07, affecting TokenHub, the ADP agent platform, and CloudBase, with migration to Hy4 (hy4) recommended. It has a hard deadline, a replacement, and even auto-switch conditions - and we happen to sell this exact model.

Plaza search for hy4 shows only one hy4-preview, priced 0.3299/M in and 0.9898/M out, labeled 95% off, 1 verified seller
Plaza search for "hy4" returns only one hy4-preview (in ¥0.3299, out ¥0.9898 per M tokens, labeled 95% off); the official migration target Hy4 isn't found

What the official side actually said

The notice also says "new and old models may produce different output; you may manually switch to another model" - auto-switch only keeps service alive, it doesn't keep output identical.

What this changes

What changes is the preview model's lifecycle. It served as the formal version's advance scout, and many teams validated or even ran production on it, since price and capability were close to GA (same $0.834 in / $2.501 out per M tokens, cache-hit $0.042). Now GA is on the shelf and the preview window closes. This isn't a price cut or an upgrade - it's a replacement you must act on; the cost of inaction isn't "a bit pricier," it's failed calls.

What it means for our users

Bottom line: if you're still on hy4-preview, it may stop working anytime after Dec 7; and we currently don't have the officially recommended Hy4.

The basis is a real test: plaza search for "hy4" returns only one hy4-preview, no hy4; the API's in-stock list has only hy3 and hy4-preview in the Hunyuan family. In other words, you can buy the preview but not the GA it wants you to migrate to.

The second layer is supply stays or goes beyond our control: hy4-preview is supplied by 1 verified seller, status 100%, throughput 36.3 t/s; whether their upstream passes through TokenHub we can't confirm for the seller - if it does, this listing drops after 12-07; if it goes through Hunyuan's native API or a third-party aggregator, it may continue. Treat it as a line that could cut anytime.

Self-made: three key dates for Hy4 preview retirement (Nov 9 announced, Dec 7 stop, Dec 31 Hy4 half-off ends) and a per-M-token price comparison, our 0.3299 vs official half-off 3 yuan, about 9.1x
Self-made comparison: three key dates and the per-M-token price math (scope noted at the end)

The money: how much pricier is migration

Per the official price page (updated 2026-11-04), the math is spelled out so anyone can re-verify:

At the CNY-base half-off (¥3.00 / ¥9.00), input cost is about 9.1x now (3.00 ÷ 0.3299 ≈ 9.09), output likewise 9.1x (9.00 ÷ 0.9898 ≈ 9.09); by the official USD list price (¥2.79 / ¥8.38) it's about 8.5x. The two differ only in base, same order of magnitude. After the half-off ends on Jan 1, 2027, it's about 18x.

Mind the scope here: ¥0.3299 is the extreme low tier of a seller's discounted quota, not the market norm; the official list price is a sticker - the two were never the same thing. The "9.1x" in the image uses the CNY-base tier, and the body lists both so you don't see only half. But the conclusion stands - you're currently saving through extreme idle-quota discounts; moving to official pays all of that back at once. For annual budgeting, use the post-promo list price.

What to do

  1. Move the model name into config today: change hardcoded hy4-preview into a config item. Ten minutes now; on 12-07 it's one line, no release.
  2. Run a comparison this week: pick 20-50 real requests (tool calls and long output first), run both models, watch three things - output format changes (JSON fields, Markdown hierarchy), whether tool-call args stay valid, and whether long context fills up. The official side itself said output may differ.
  3. Add a "model unavailable" alert: track hy4-preview's 4xx/5xx separately; without monitoring, probe at least hourly.
  4. Buying here: check the seller on the detail page and ask directly "still supplying after Dec 7?"; meanwhile set a backup - the sibling previous-gen hy3 is still on sale and far cheaper (in ~¥0.05, out ~¥0.22), but it's a generation behind, only as a degradation fallback. Don't keep just one line.
  5. Going official: there's a half-off window through 12-31 to migrate and cut cost; but it reverts to full price on 2027-01-01, so budget at full price.

One line

The preview's "temporary" has really expired this time. The notice has no new capability and no price cut; its whole value is in the date: 2026-12-07 00:00. From the Nov 9 announcement to shutdown is under a month - just finish the swap before Dec 7; past that day it's not "a bit pricier," it's "calls fail."

(Source: Tencent Cloud announcement 101604 (2026-11-09); prices from TokenHub "Model Pricing" (updated 2026-11-04); our prices are the model plaza lowest quote at 2026-11-10 09:40, moving with seller listings. The two sides use different scopes; the multiple is only an order-of-magnitude reference, not a price promise.)

相关文章

GPT-6.2 Cuts Prices 20%: What the AI Price War Means

GPT-6.2 cuts list prices 20%, yet platform discounts run deeper: gpt-6-astra trades at under 1% of list. What this week's AI price war means for buyer

What Is Prompt Caching? How Cache Tokens Cut API Bills

Prompt caching stores a repeated prompt prefix and bills cache reads at 5-12% of normal input. Real prices, hit conditions, and how to cut your API bi

All 12 Claude Models Compared: Price and Best Use

Twelve Claude versions are on sale. The newest opus-5-5 costs $8.91 a month; the oldest opus-4-6 costs $49.50. Full table and a pick for each workload

Which AI Model Writes Code Best? Speed and Price Tested

gpt-5.6-terra and deepseek-v4.1-flash win on price and reliability; gemini-3.8-flash is fastest. Effective prices, measured speed, and a three-tier pi

One API Key, Every Model: How to Integrate a Multi-Model Marketplace

Point your base URL at the platform and one key covers all 138 on-sale models across 51 vendors, in both OpenAI and Anthropic request formats. Copy-pa