November 5, 2026 | AI | By Mahesh

Tencent's Hy4 costs $0.834 per million input tokens and $2.501 per million output tokens, and an Artificial Analysis score of 35 puts it level with Grok 4.7 at roughly a tenth of that model's cost, according to AlphaSignal [1][2]. Tencent announced the official Hy4 on November 4, 2026 and kept API pricing at the preview level [3]. Tencent's Hong Kong shares closed 7.17% higher on November 5, though the sources do not link the move to Hy4 alone [3].

What changed with the official release

BigGo Finance reports that Hy4 grew from under 300 billion parameters in the Hy3 series to 770 billion, added visual understanding as a first for the series and added command line and graphical computer operation [3]. The parameter count comes from Hong Kong media reports, and Tencent plans to release the weights without a date and to launch a lower-cost Hy4 flash version soon [3]. Tencent names office collaboration, financial analysis, front-end development and game development as target workloads [2].

Holding the price steady is a notable choice, since eesel.ai notes that Hy4 costs about five to six times more than Hy3, so teams moving from the older model face a real increase [1]. BigGo's view is that stable pricing could speed enterprise and consumer adoption, and that the share gain mostly reflects hopes for wider AI investment at Tencent [3]. The article cites no named analysts.

Under the hood: size and hardware cost

Hy4 is a 770-billion-parameter mixture-of-experts model with 49 billion active per token, but all weights must sit in GPU memory [1]. The full BF16 weights are about 1,848 GB, MXFP8 weights are about 924 GB and a Q4 GGUF is around 450 GB, so roughly 512 GB is the realistic minimum for local use [1]. An 8x H100 node with 640 GB cannot run it, and eesel.ai estimates that renting an 8x B200 node costs tens of thousands of dollars per month [1]. That estimate is the site's own. For another model with a 1M context window, see Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks.

Independent scores and what they show

AlphaSignal reports Hy4 at 35 on the Artificial Analysis Intelligence Index v5, level with Grok 4.7 at its highest effort setting and one point above Qwen3.8 Max at 34 [2]. Claude Haiku 5.5 scores 36 at maximum effort and costs about half as much as Hy4, by the same article [2]. Hy4 also scores 31% on FrontierCode 2.0, 37% on Terminal-Bench 4.0 and 33% on HLE-Diamond [2].

One number cuts against the price story. Hy4 posts 50% accuracy on AA-Omniscience but an 89% hallucination rate, which means it often answers when it should decline [2]. The independent Terminal-Bench 4.0 score of 37% is also lower than the 44.4 that Tencent reports itself, so the two figures should not be mixed [2]. AlphaSignal's page gives no calendar date, so confirm the figures against Artificial Analysis before quoting them.

API pricing against rivals

Hy4 charges $0.834 for input, $2.501 for output and $0.042 for cached input per million tokens, and there is no permanent free API tier [1]. Tencent's WorkBuddy and CodeBuddy offered a two-week free window at launch [1]. The table below uses eesel.ai's comparison, which was dated August 30, 2026 and lists Kimi K3 and DeepSeek V4 and not the newer K3.1 or V4.1 releases, so treat rival prices as a snapshot.

Hy4 is cheaper than Kimi K3, Qwen 3.8 Max and GLM-5.3, and second cheapest on output behind DeepSeek V4 Flash, by eesel.ai's table [1]. Output is the meter to watch, since Hy4 tends to generate many tokens, and cache hits can cut input cost for workloads that repeat large prompts [1].

Limitations and who should care

Hy4 over-verifies its work, and the extra output tokens can push real costs above the listed rate [1]. DeepSeek V4 Flash is far cheaper on both meters, and one Reddit user quoted by eesel.ai argued that cheaper models can perform similarly on many tasks [1]. The full-release weights are not out, the parameter count rests on media reports and the share move is a market reaction and not a performance measure [3].

For most teams, the practical choice is hosted access through Tencent or a router, with the 89% hallucination rate in mind for any task that needs refusals. Run a sample of your own prompts against Hy4 and DeepSeek V4 Flash and compare cost per correct answer. For an open-weight alternative, see Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing.

Sources

1. eesel.ai. "Tencent Hy4 pricing 2026: API rates, hardware cost, rivals compared" (dated August 30, 2026). eesel.ai
2. AlphaSignal. "Tencent's Hy4 Brings 1M Context and Agent Tools at a Tenth the Cost" (undated page). alphasignal.ai
3. BigGo Finance. "Tencent Shares Jump as Much as 6% After Official Release of 770 Billion-Parameter Hy4." finance.biggo.com

Read More on Depth Grid

• Nvidia in Talks to Deepen Stake in or Buy Reflection AI
• Claude Haiku 5.5 Launch: $0.10 Input, 1M Context, Benchmarks
• Meta Muse Wonder Launch: 1M Context, $4 API Price, Benchmarks
• Mistral Large 4 Preview: 1.05T Open-Weight MoE, Specs and Pricing
• Tencent Hy4: 770B MoE, 1M Context, 6 Yuan per Million Tokens

Article by Mahesh | Depth Grid