Hy4 Flash API Pricing: What Tencent's Model Actually Costs
Hy4 Flash API pricing explained: per-token rates, the half-price Cloud Token Hub promo through year-end, and how it stacks up on cost per task.

How much does the Hy4 API cost per million tokens?
Tencent prices the Hy4 API at roughly 83 cents per million input tokens and about $2.50 per million output tokens, the same rate that applied to the earlier Hy4 Preview build. Through Tencent’s Cloud Token Hub, that rate is currently cut in half, running through December 31st. Hy4 Flash, the lighter variant built for everyday productivity work, is positioned as the cost-efficient option in the lineup, with Tencent citing cost-per-task figures far below both Hy4 and rival models rather than a separate published per-token rate.
TL;DR
- Hy4’s standard API pricing sits around 83 cents per million input tokens and $2.50 per million output tokens, unchanged from the Hy4 Preview rate.
- Cloud Token Hub currently bills Hy4 usage at half price, a promotion that runs through December 31st.
- Hy4 Flash isn’t the same SKU as Hy4. It’s a separate, smaller-footprint model (220B total parameters, 12B active per token) that Tencent built specifically to cut cost per task rather than cost per token.
- On Tencent’s internal WorkIndex benchmark, Hy4 Flash’s average cost per task comes in at about 1/13th of Hy4 Preview, 1/20th of full Hy4, and roughly 1/8th of DeepSeek-V4.1-Flash on its Max tier at peak pricing.
- A blind user evaluation put Hy4 Flash’s cost at RMB 0.26 per task, with a median completion time of 6.7 minutes, while user satisfaction landed close to Claude Haiku 5.5.
- Hy4 Flash ships as open weights (standard, FP8-quantized, and base versions), so teams can also self-host instead of paying per-token API fees at all.
- Tencent recommends Hy4 Flash for routine work like document processing, slide creation, simple frontend builds, and computer-use tasks, reserving full Hy4 for long-horizon software engineering where the per-token cost matters less than raw capability.
What is the actual per-token price for Hy4?
The headline numbers, confirmed in Tencent’s own materials, are straightforward: about $0.83 per million input tokens and about $2.50 per million output tokens. That’s the same pricing Tencent used for Hy4 Preview, so the jump from preview to the official November 4th release didn’t come with a price hike. For teams already budgeting around Hy4 Preview costs, nothing changes on that front.
What does change, temporarily, is the discount layered on top. Tencent’s Cloud Token Hub is billing Hy4 usage at half price, and that promotion is scheduled to run through the end of the year. For any team doing meaningful volume, that’s effectively a 50% reduction on both input and output token costs during the promotional window, which makes now a reasonable time to run larger evaluation batches or migrate workloads before the discount potentially expires.
Why does Hy4 Flash use a different cost model?
Hy4 Flash isn’t a cheaper tier of the same model. It’s an architecturally distinct release: 220 billion total parameters (including a 63.5B “Engram” component for n-gram embeddings) but only about 12 billion active parameters per token (11.3B text plus 0.43B vision). That active-parameter count is what keeps inference cheap, since it determines how much compute actually fires on each forward pass, not the total parameter count sitting on disk.
Tencent’s documentation frames Hy4 Flash’s value proposition around cost per task rather than cost per token, because a model can be cheap per token and still expensive overall if it needs more tokens, more turns, or more retries to finish a job. The company says reinforcement learning during post-training cut average end-to-end output tokens by 50% compared to earlier models, and architectural changes (hybrid sparse plus sliding-window attention, a technique called LatentMoE, and the Engram n-gram lookup system) shrank the KV cache to roughly 1/24th the size used by Hy3 and 1/6th the size used by full Hy4. A smaller KV cache means less memory pressure during inference, which translates into higher throughput and lower serving cost at scale. Tencent also reports decoding speeds above 150 tokens per second using 4-bit quantization and a speculative decoding module called DFly2.
Is Hy4 Flash actually cheaper in practice?
By Tencent’s own benchmark, yes, by a wide margin. On an internal benchmark called WorkIndex, built around typical WorkBuddy-style productivity tasks (slide creation and editing, document editing, spreadsheet repair, audio/video processing, visual understanding), Hy4 Flash scored 42.1, slightly ahead of Hy4 Preview’s 39.7. But the cost difference is the bigger story: Tencent reports Hy4 Flash’s average cost per task is about 1/13th of Hy4 Preview, 1/20th of full Hy4, and roughly 1/8th of DeepSeek-V4.1-Flash-1017 when that competitor is measured on its Max tier at peak-hour pricing.
Build an enterprise app with Dmitry Shapiro. Leave certified.
A separate blind evaluation involving 85 internal users across 259 tasks in 27 categories backs this up with real usage data rather than synthetic benchmarks. Hy4 Flash scored 2.79 out of 4 on user satisfaction, just behind Claude Haiku 5.5’s 2.81, while finishing tasks in a median of 6.7 minutes at a cost of RMB 0.26 per task, the lowest cost and fastest completion time among the models tested in that round.
For teams doing high-volume, repetitive work (document QA, chart reading, spreadsheet fixes, simple web builds), that cost-per-task advantage compounds fast. Tencent itself recommends Hy4 Flash for exactly those categories: deep search, document processing, slide creation, simple frontend demos, and computer-use tasks. For long-horizon software engineering work where correctness over many steps matters more than speed or price, the company still points users toward full Hy4, which scored 51.4 on the same WorkIndex benchmark.
How does the Cloud Token Hub discount work?
Cloud Token Hub is the billing mechanism Tencent uses to meter and discount API usage, and right now it’s applying a 50% reduction to Hy4 pricing through December 31st. That brings the effective rate to roughly 41.5 cents per million input tokens and $1.25 per million output tokens for the duration of the promotion. There’s no indication in Tencent’s materials that this discount extends automatically into the new year, so anyone building cost projections around current rates should treat the discounted price as temporary and budget for the standard rate afterward.
It’s also worth separating the API-pricing conversation from the open-weight release. Hy4 Flash, along with an FP8-quantized version and a pre-trained base model, is available as open weights on Hugging Face, ModelScope, and GitCode. That means organizations with their own GPU infrastructure can skip per-token API billing entirely and run inference costs based purely on hardware and electricity, an option that matters most for teams with sustained, high-volume workloads where self-hosting amortizes below API pricing over time.
Frequently Asked Questions
How much does the Hy4 API cost per million tokens?
Roughly 83 cents per million input tokens and about $2.50 per million output tokens, the same rate as Hy4 Preview. Cloud Token Hub currently halves that rate through December 31st.
Is there a separate price for Hy4 Flash?
Tencent hasn’t published a distinct per-token rate for Hy4 Flash in its API pricing materials. Instead, the company emphasizes cost per task, reporting figures roughly 1/13th of Hy4 Preview and 1/20th of full Hy4 on its internal WorkIndex benchmark, and RMB 0.26 per task in a blind user evaluation.
Does the Cloud Token Hub discount apply to both input and output tokens?
Based on Tencent’s pricing statement, the half-price promotion applies to the standard Hy4 API rate overall, covering both input and output token costs during the discount window through year-end.
Can I avoid API pricing entirely by self-hosting Hy4 Flash?
Yes. Hy4 Flash, an FP8-quantized version, and a pre-trained base model are released as open weights on Hugging Face, ModelScope, and GitCode, so teams with their own infrastructure can run it without per-token API billing.
Is Hy4 Flash cheaper than DeepSeek or Claude alternatives?
Tencent’s own benchmarks claim Hy4 Flash costs about 1/8th of DeepSeek-V4.1-Flash-1017 at Max-tier peak pricing, and a blind evaluation showed it costing less per task than the other models tested while reaching user satisfaction close to Claude Haiku 5.5. These are Tencent-reported figures, not independently audited numbers.