Tencent Hunyuan Hy4 Goes GA: Same List Price, a 50% Timer, and Vision + Computer Use

The facts, graded (Nov 4)

• The rate card, verified by direct read. Tencent Cloud's TokenHub price document lists hy4 at ¥6 input / ¥18 output / ¥0.3 cache hit per 1M tokens (Guangzhou, pay-as-you-go), with hy4-preview priced identically. A limited-time note states Hy4 calls settle at 50% of list until 2026-12-31 23:59:59 Beijing time — effective ¥3/¥9, cache ¥0.15. The batch-task tier still lists preview only (¥3/¥9/¥0.15).
• The international channel, verified by direct read. OpenRouter carries tencent/hy4 at $0.417 in / $1.251 out per 1M, alongside tencent/hy4-preview. Those USD figures correspond to the discounted ¥3/¥9 at recent CNY/USD rates — not the ¥6/¥18 list — so international builders are getting the launch discount through the gateway. Our read of the parity, based on list prices published by each platform; OpenRouter's own page does not state the ¥ basis.
• What GA adds (vendor-reported). Vision understanding and computer use join the flagship: Tencent reports OSWorld 2.1 at 78.9 (full/partial), versus Kimi K3.1 Pro at 79.6 and Claude Opus 5.5 at 81.8 in the same release table, with dedicated training for professional software (Blender, CAD) and cross-application long-horizon tasks.
• Preview→GA long-horizon jumps (vendor-reported). DeepSWE v1.1 64.3→74.9; Terminal-Bench 4.0 26.3→44.4; APEX-Agents v1.1 (pass@1) 53.9→75.2; AutomationBench v1.0.6 36.3→56.2. FrontierCode 2.0: 31.0 for Hy4 vs DeepSeek-V4.1-Flash-1017 28.5, Kimi K3.1 Pro 25.5, GLM-5.3 21.6.
• Finance and search (vendor-reported). A 213-expert crowdtest over 579 real financial tasks scores Hy4 at 3.04 vs DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro and GLM-5.3 at 2.68. In Tencent's in-house search test, hallucination rate fell 15.76%→5.45% versus preview.
• The open-weights question. The August preview was released open under Apache 2.0 (770B total / 49B activated per token, 1M context per the preview documentation). For GA, Tencent says weights are planned open, with vLLM/SGLang/Inferact ecosystem support, a system card and a fuller safety evaluation promised at release — but no GA repository existed as of Nov 4, 2026. Treat today's downloadable weights as preview, not GA.
• Ecosystem signals to watch. Hy4 Flash (lower cost, faster) is announced as coming soon; it is not on the TokenHub price page as of Nov 4. Hunyuan's research channel also moved the same week — the Hyra-results repository (updated Nov 3) ships the 16M-parameter HyCell single-cell model and HyPDE — a signal the lab is actively shipping around the flagship.

Verification on Nov 4, 2026: Tencent Cloud TokenHub price document and OpenRouter model pages read directly; preview specification from Tencent's Aug 28 newsroom release and GitHub repository; GA capability figures are Tencent's release-note claims relayed by IT Home, Gelonghui and the NLPer writeup — vendor-reported, no independent benchmark reproduction performed.

Context: where ¥3/¥9 lands on the ladder

On Tencent Cloud's own bill, the promotional ¥3/¥9 puts Hy4 under GLM-5.3 (¥8/¥28, cache ¥2), DeepSeek-V4-Pro (¥12/¥24, cache ¥1) and Kimi K3.1 Pro (¥13/¥68, cache ¥1.3), while DeepSeek-V4.1-Flash stays cheaper still (¥1/¥4 off-peak, ¥2/¥8 peak). At list ¥6/¥18, Hy4 remains one of the cheaper frontier flagships on the board — the discount is what makes it aggressive for the next two months.

Two structural notes for buyers. First, the timer matters: the 50% settlement discount runs through Dec 31, 2026, so any budget math built on ¥3/¥9 should carry an expiry date. Second, cache pricing is unusually deep: ¥0.3 per 1M cached input (¥0.15 at promo) is 20x cheaper than uncached input at list — for agent workloads with long, stable prefixes, that ratio is the real price story.

This is also ChinaModelAPI's first Hunyuan brief — the flagship's GA closes a category gap on a site that has tracked DeepSeek, GLM, Kimi and Qwen pricing all year. The GA-day coverage of peers is directly comparable: GLM-5.3's one-day slip and GA, DeepSeek-V4.1-Flash's official launch, and the flash-tier price war that frames the low end.

What this means for builders

The practical protocol on GA day: confirm the exact model id (hy4 on TokenHub; tencent/hy4 on OpenRouter) in your provider dashboard before production traffic, and if the model fits your workload, size commitments against the Dec 31 discount deadline rather than the list price. International billing through OpenRouter at $0.417/$1.251 is currently cheaper than paying the ¥ list through a card — one more reason the promo math follows the workload.

Benchmark discipline still applies: every capability number in this brief is Tencent's own, and the preview weights are not the GA model. Run one real request and read the usage line; for self-host comparisons, benchmark the preview weights against your serving stack while waiting for the GA release. For agent workloads, model the cache math — ¥0.3/1M cached input at list (¥0.15 at promo) changes routing economics for long-prefix agents.

Watch four signals for the follow-up: the GA weights release (with system card and safety evaluation), the Hy4 Flash rate card, an Artificial Analysis score for Hy4 GA, and any change to the 50% discount line. Any of them moves, and this page gets an update entry.

Access from outside China — and the usual disclaimer

This site tracks Chinese model APIs for international builders: official endpoints, pricing and open-weights status. Hy4 sits in the same arena as the models in the comparison table and the cheapest-API guide; the international channel today is OpenRouter.

Independence disclaimer: ChinaModelAPI is an independent information site, not affiliated with Tencent, Hunyuan or any model vendor. Rate-card and listing facts are from direct reads of the TokenHub document and OpenRouter; all capability figures are Tencent's own, relayed with attribution; nothing here is a routing, availability or pricing promise.

Primary sources

• Tencent Cloud TokenHub — model price document (hy4 ¥6/¥18/¥0.3, 50% note to Dec 31, 2026; read directly)
• OpenRouter — tencent/hy4 listing ($0.417/$1.251 per 1M; read directly)
• Tencent newsroom (Aug 28, 2026) — Hy4 preview release: 770B/49B MoE, 1M context, Apache 2.0 open weights
• IT Home (Nov 4, 2026) — Hy4 GA report: vision, computer use, strongest-open-tier claim
• NLPer (WeChat, Nov 2026) — the release-note walkthrough this brief cross-checked; benchmark tables relayed
• GitHub — Tencent-Hunyuan/Hyra-results (research artifacts; updated Nov 3, 2026)
• 站内对照：GLM-5.3 GA · DeepSeek-V4.1-Flash 官宣 · 最低价指南 · 对比表

FAQ (2026)