Tencent Quadruples HY2.0 Price as Agent Costs Mount
Key insights
- Tencent raised its HY2.0 Instruct model prices up to fourfold since late March 2026 as agentic workloads drove infrastructure costs higher.
- ByteDance hiked Seedance generative video subscriptions three times in April alone, with the annual fee rising from 2,599 to 3,099 yuan.
- DeepSeek launched V4 Pro on April 24 at $3.48 per million output tokens, a fraction of OpenAI GPT-5.5's $30 per million rate.
Why this matters
The 10-to-100x infrastructure cost multiplier from agentic AI use is now an economic constraint playing out in live pricing decisions, meaning any product priced for light query traffic faces forced repricing as user behavior shifts to agents. For founders designing AI-native workflows, this changes build-versus-buy calculus: hardware and model efficiency are increasingly required to sustain competitive pricing, as DeepSeek's dependency on Huawei's Ascend 950 chips makes explicit. The repricing by Tencent, Baidu, Alibaba, and ByteDance dismantles the structural price-advantage narrative Chinese labs used against OpenAI and Anthropic, opening space for Western providers to compete on workflow value rather than sticker cost.
Summary
China's largest AI labs are walking back the price-war logic that built their global cost case. Tencent, Baidu, Alibaba, and ByteDance all raised prices since late March 2026, with Tencent lifting its HY2.0 Instruct model as much as fourfold.
The culprit is agentic AI: moving from simple queries to coding and research tasks can push per-user infrastructure costs 10 to 100 times higher, breaking the unit economics of below-cost pricing.
Essentially: (Tencent, Baidu, ByteDance, Alibaba) found that heavy AI usage intensifies costs rather than diluting them.
- ByteDance's Seedance annual fee rose from 2,599 yuan to 3,099 yuan ($371 to $442) across three April hikes.
- DeepSeek holds at $3.48 per million tokens; OpenAI GPT-5.5 charges $30.
- Huawei's Ascend 950 chips, expected at scale in H2 2026, are DeepSeek's stated bet on sustaining that gap.
The repricing puts direct pressure on the claim that Chinese AI is structurally cheaper than Western rivals.
Potential risks and opportunities
Risks
- DeepSeek faces margin erosion if agentic workload demand grows faster than Huawei Ascend 950 chip efficiency gains materialize in H2 2026.
- Tencent and ByteDance risk enterprise customer defections if sequential price hikes outpace demonstrable workflow ROI over the next two quarters.
- Anthropic, which already closed one pricing loophole tied to third-party tool integrations like OpenClaw, faces ongoing pressure to identify and close similar exploits as agentic use cases proliferate.
Opportunities
- Huawei gains strategic leverage with Chinese AI labs as its Ascend 950 chips become the deciding hardware variable for which providers can sustain low inference pricing past H2 2026.
- Enterprise consultancies specializing in agentic workflow optimization gain a clear cost-reduction pitch to Chinese enterprise clients now confronting 10-to-100x infrastructure cost increases.
- Western AI providers including Anthropic and OpenAI gain a repricing story for enterprise sales conversations where Chinese labs previously held an unambiguous price advantage.
What we don't know yet
- Whether DeepSeek's $3.48/million token pricing remains viable once Huawei's Ascend 950 chips ship at scale in H2 2026 and real inference costs at agentic load become measurable.
- Which specific enterprise customer segments or workflow types are generating the 10-to-100x cost multiples cited as the article's central data point.
- Whether Alibaba's daily-batched, request-capped subscription model (90,000 requests per month at 200 yuan) faces the same upward repricing pressure as flat-rate competitors already have.
Originally reported by thewirechina.com
Read the original article →Original headline: China's AI Firms Abandon Price War as Agentic Workloads Cost 10–100× More — Tencent Raises HY2.0 by 4×, DeepSeek Holds Out