Tencent on Monday launched Hy4 flash, a lower-cost AI model built for the kind of work most people actually do with chatbots: research, slide creation and spreadsheet editing. The release is the newest member of the company's Hy4 family, and it arrives with an aggressive pitch — on Tencent's internal WorkIndex benchmark, Hy4 flash cost one-twentieth as much per task as the full Hy4 model, according to TechNode reporting on the launch.

The move is the latest salvo in China's price war over everyday AI usage, a day after rival Alibaba launched its flagship Qwen4-Max with its own cost-cutting claims. For a running view of how fast model economics are shifting, follow our latest AI developments.

Pricing built for high-volume work

Tencent published list prices that are unusually low even by the standards of China's discount-heavy API market. Hy4 flash costs $0.07 per million input tokens, $0.281 per million output tokens, and just $0.003 per million cached input tokens — a structure that rewards applications which reuse the same context across many requests, according to TechNode.

The headline comparison comes from Tencent's own WorkIndex benchmark, which measures the cost of completing realistic office tasks rather than raw token prices. On that measure, Hy4 flash came in at one-twentieth of the cost per task of the full Hy4 model. Tencent is also running launch promotions: WorkBuddy users can access the model for free through November 30, and Tencent Cloud's TokenHub API platform is offering access at half the list price.

Those numbers are vendor-reported, and internal benchmarks are designed to flatter the products that run on them. But the direction is consistent with what Tencent's rivals are doing: Alibaba's Qwen4-Max launch on Sunday leaned on the same pitch of dramatically lower task costs, and third-party trackers have noted both companies launching flagship models within a day of each other.

Small active footprint, long memory

Under the hood, Hy4 flash is a 224-billion-parameter model that activates only 12 billion parameters per forward pass — a mixture-of-experts-style design that keeps inference cheap while retaining a large total capacity. The model supports multimodal inputs and a context window of up to one million tokens, TechNode reported.

That combination matters for the office-work use case Tencent is targeting. A million tokens is enough to hold long documents, spreadsheets or research material in a single session without chunking, and a small active parameter count means the model can serve those long sessions at a fraction of the compute cost of a dense model with similar headline size. The cached-input pricing reinforces the same design goal: applications that keep a large, stable context and send small updates pay almost nothing for the repeated portion.

How it scored in Tencent's blind test

Alongside the benchmark numbers, Tencent pointed to a separate blind evaluation in which 85 WorkBuddy users rated model outputs across 259 tasks. In that test, Tencent said Hy4 flash scored close to Claude Haiku 5.5 — Anthropic's small, fast workhorse model — in user satisfaction, while recording the lowest cost and the shortest completion time among the models tested, per TechNode.

The comparison set is telling. Rather than benchmarking against frontier giants, Tencent positioned Hy4 flash directly against the class of models that actually handle high-volume, cost-sensitive workloads: fast, inexpensive models tuned for everyday tasks. Claiming parity in satisfaction with an established leader in that class — while undercutting it on both speed and price — is the specific gap the model is built to exploit.

Independent verification will take time. User-satisfaction studies run by the vendor, like internal benchmarks, deserve skepticism until third parties replicate them. But the pricing is public and checkable from day one, and it is the part most likely to move developer behavior.

Where it runs, and what happens next

Hy4 flash is available immediately in Tencent's WorkBuddy assistant and in Yuanbao, the company's consumer chatbot, as well as through Tencent Cloud's TokenHub API service. Developers outside Tencent's ecosystem will get access too: the company said it plans to release the model's weights and make it available on OpenRouter, the third-party model aggregator, according to TechNode.

A planned weights release is consistent with how Tencent has handled the rest of the Hy4 family, parts of which have been promised to the open-source community. It also distinguishes Tencent from Alibaba's Qwen4-Max launch this week, which coverage has described as keeping the flagship's weights private. If Tencent follows through, Hy4 flash would give developers a genuinely open option in the budget multimodal tier — and put pressure on every other lab selling cheap office-work models.

What to watch

Three things will determine whether Hy4 flash's launch translates into real adoption. First, whether the weights release actually ships, and under what license. Second, whether independent users confirm the quality claim against Claude Haiku 5.5 in real workloads rather than curated test sets. Third, whether the promotional pricing — free in WorkBuddy through November 30, half-price API access on TokenHub — converts into list-price usage once the discounts expire.

For now, the message from China's two largest consumer internet companies is the same: the fight over AI has moved from raw capability to the cost of ordinary work, and both Alibaba and Tencent are willing to cut prices by orders of magnitude to win it.

---

Stay Ahead of AI

Get the latest AI news, analysis, and breakthroughs — all in one place.

Read more AI news →