Tencent Hy4 Hits TokenHub and OpenRouter—Vision In, 50% Off Through Year-End

Tencent Cloud's TokenHub model list and pricing pages both show a last-updated stamp of 2026-11-04 12:13:30. The language-model table now includes formal Hy4 with call id hy4. Relative to Hy4 preview (hy4-preview) on the same page, the formal row adds image understanding, while keeping deep thinking (retained-style), structured output, Function Calling, Cache, a 1M context window, 960k max input, and 64k max output. OpenRouter listed tencent/hy4 almost in lockstep—API created=1793766031 resolves to 2026-11-04 12:20:31 Asia/Shanghai—with modality text+image→text and a page label of 50% off. DigIntoAI has not called Hy4 on TokenHub or OpenRouter; figures below come from TokenHub documentation and the OpenRouter model / API self-report.

Related DigIntoAI notes: Tencent Yuanbao Search WSA Pricing (October 2026), Hunyuan Opens Hy-Nomas: Interactive Video World Model on a 4090, Tencent Opens Hy-SimulMT — Download Isn't Instant Interpreting, and OpenAI API Pricing (November 2026).

Formal Hy4 vs Hy4 preview on TokenHub

TokenHub's model list is the clearest side-by-side:

So the capability gap on the list page is picture / image understanding on formal Hy4, not a new context size. TokenHub's capability glossary groups "image and video understanding" as a platform concept, but the Hy4 row itself only ticks image understanding—so video understanding is not listed as a Hy4 capability.

OpenRouter's model description for tencent/hy4 frames the same gap in English: the formal release is cast as a broad upgrade over Hy4 Preview, with stronger agentic and long-horizon productivity framing, and "first" introduction of visual understanding. FAQ-style notes on that page also say tools / tool_choice and JSON-schema response_format are supported. Those are OpenRouter page claims; DigIntoAI has not verified tool loops or vision quality.

TokenHub list price and the year-end 50% window

On the TokenHub pricing page (same 2026-11-04 12:13:30 stamp), online inference for Guangzhou postpaid token billing lists:

A callout above the tables states the Hy4 limited-time discount: from the notice's publication through 2026-12-31 23:59:59 (Beijing time), Hy4 calls settle at 50% of list price; the discount applies only to usage actually incurred in that window, then list rates resume. Applying that 50% rule to the 6 / 18 list line yields about 3 / 9 RMB per million tokens (input / output)—a document-based estimate, not a checked invoice. DigIntoAI has not reconciled a bill.

The batch pricing table currently visible only shows Hy4 preview at 3 / 9 / 0.15. No matching batch row for formal hy4 appears there.

OpenRouter mirror: price, modality, timing

OpenRouter API entry id=tencent/hy4 (canonical slug tencent/hy4-20261104) reports:

• created=1793766031 → 2026-11-04 12:20:31 Asia/Shanghai
• context_length=1048576 (1M); top_provider.max_completion_tokens=65536
• Pricing: prompt 0.000000417, completion 0.0000012505, input_cache_read 0.000000021 → about $0.417 / $1.251 / $0.021 per million tokens
• Architecture: modality text+image→text; inputs text + image; output text
• Supported parameters include tools, tool_choice, response_format / structured outputs, and reasoning-related flags (as listed by the API)

The model page labels the current rate as 50% off, with struck list-style figures around $0.834 / $2.501—consistent with half of those list-like numbers. DigIntoAI has not confirmed how long OpenRouter keeps that promo, or whether TokenHub's BJT year-end window and OpenRouter's discount end on the same clock.

For contrast, tencent/hy4-preview on OpenRouter is older (created ≈ 2026-08-28 14:09:35 Asia/Shanghai), modality text→text (no image), and sits near $0.834 / $2.501 list-style pricing. That preview path is the text-only predecessor; formal hy4 is the vision-capable SKU in the current API lineup.

Weights: preview on Hugging Face; formal not verified public

A Hugging Face / search check still surfaces public tencent/Hy4-preview (and FP8) weights under Apache-2.0, with preview self-descriptions around ~770B total / ~49B active and 1M context. Formal tencent/Hy4 / tencent/hy4 was not verified as a public weight drop—so the formal release is not described here as open-sourced or already on HF. The anchor here remains API availability on TokenHub and OpenRouter.

What DigIntoAI has not done

DigIntoAI has not measured TokenHub or OpenRouter latency, throughput, image-understanding quality, agent / tool chains, or discounted invoices for Hy4. OpenRouter page throughput or latency numbers—if any—are page self-reports only. Product docs and promo windows can change; TokenHub documentation and the live OpenRouter model / API entry take precedence.

Quick facts

1. Who. Tencent / Hunyuan via TokenHub; also OpenRouter tencent/hy4.
2. What. Formal Hy4 (hy4) with image understanding, 1M / 960k / 64k, deep thinking / structured output / FC / Cache.
3. When. TokenHub pages 2026-11-04 12:13:30; OpenRouter created 2026-11-04 12:20:31 Asia/Shanghai.
4. TokenHub price. List 6 / 18 / 0.3 RMB/M; 50% through 2026-12-31 23:59:59 BJT → ~3 / 9 estimate; batch table currently = preview only.
5. OpenRouter price. ~$0.417 / $1.251 (cache ~$0.021), 50% off, text+image→text.
6. Not claimed. Formal open weights; video understanding; permanent half-price; DigIntoAI benchmarks.
7. Related. Yuanbao WSA pricing, Hy-Nomas, Hy-SimulMT, OpenAI API pricing.

In short

Formal Hy4 is on TokenHub and OpenRouter as of 2026-11-04 Asia/Shanghai, with image input versus text-only Hy4 preview, shared 1M-class context, list rates of 6 / 18 RMB/M on TokenHub under a documented 50% window through year-end BJT, and OpenRouter rates near $0.417 / $1.251 under a 50% off label. DigIntoAI has not benchmarked the model; the numbers above follow TokenHub documentation and OpenRouter's model / API self-report.