Anyone weighing a cheaper AI provider is really weighing two questions: how much the switch saves, and what legal jurisdiction their prompts land in. China's AI industry just sharpened both. DeepSeek's website logged 541 million monthly visits in May 2026, up 11.22 percent month over month, keeping it first in China and fifth globally among all AI products, according to traffic figures reported by Pandaily. Days earlier, the company locked in API prices at one-quarter of its April list rates. For US readers, the stakes are concrete: pressure on what OpenAI, Anthropic, and Google charge, and a data-jurisdiction tradeoff the headline prices leave out.
How Big Is DeepSeek's Lead in China's AI Traffic Race?
The May traffic table shows a market with one giant and a long tail. Behind DeepSeek's 541 million visits, Nami AI Search took second domestically with 198 million, ByteDance's Doubao logged 164 million, and Moonshot AI's Kimi and Alibaba's Qwen rounded out the top five at 46 million and 44 million visits, per the same Pandaily report. Baidu, widely written off after a quiet 2025, staged what Pandaily calls a multi-pronged comeback, with Baidu AI Search growing 20.65 percent month over month to 39.43 million visits.
The field itself has narrowed dramatically. A Q2 2026 landscape report from Digital Applied counts roughly ten serious Chinese providers left standing, down from the dozens of labs publishing competing models in 2024: Xiaomi, Alibaba, Zhipu, DeepSeek, Moonshot, MiniMax, StepFun, ByteDance, Baidu, and Tencent. Usage data shows the same: on OpenRouter, Xiaomi alone processes 4.21 trillion tokens a week for a 21.1 percent share, and six Chinese providers together exceeded 45 percent of total weekly volume in April.
Why Did DeepSeek Make Its V4-Pro Price Cut Permanent?
DeepSeek launched V4-Pro on April 24, 2026 at $1.74 per million input tokens and $3.48 per million output tokens, then immediately discounted it 75 percent in what Bloomberg framed as the latest salvo in China's AI price war. The discount was scheduled to lapse at the end of May. Instead, on May 23 the company announced it would stick: after the promotional window closed on May 31 at 15:59 UTC, V4-Pro pricing was permanently set at one-quarter of the original list, $0.435 per million input tokens and $0.87 per million output tokens, per InfoWorld and DeepSeek's own API pricing page.
Against current US rate cards, the gap is stark. OpenAI's GPT-5.2 lists at $1.75 input and $14.00 output per million tokens, per Price Per Token. Anthropic's flagship Claude Opus 4.8, released May 28, runs $5.00 and $25.00, per Anthropic's pricing documentation. Google's value-tier Gemini 3.5 Flash costs $1.50 and $9.00, as Simon Willison documented at launch. On output tokens, DeepSeek's flagship now undercuts Claude Opus 4.8 by roughly 29 to 1.
That gap rests on engineering choices: mixture-of-experts architectures that activate only a fraction of total parameters per query, subsidized compute, and margins thin enough to buy market share. And the race to the bottom is not uniform: the Digital Applied report notes Alibaba raised AI computing prices by up to 34 percent on March 18 because demand outran its infrastructure, while Zhipu AI raised API prices 83 percent earlier this year. Cheap tokens are a weapon some Chinese firms deploy and others are quietly retiring as capacity tightens.
Is Baidu Really Open-Sourcing Ernie This Month?
The second front is open source, and the record needs precision: a widely repeated version of the story is a year off. Baidu's landmark release happened on June 30, 2025, when it open-sourced the entire ERNIE 4.5 family: ten variants from a 0.3-billion-parameter dense model to a 424-billion-parameter mixture-of-experts model, all under the Apache 2.0 license, per Baidu's announcement and reporting from CNBC and TechTarget. No comparable June 2026 announcement exists.
What Baidu runs in 2026 is a split strategy. The frontier stays closed: ERNIE 5.0, the 2.4-trillion-parameter multimodal model unveiled in November 2025, is proprietary, per eWeek, and ERNIE 5.1, released May 8, 2026 and given a wide rollout at Baidu's Create conference on May 13-14, is hosted-only through Qianfan and ernie.baidu.com, per Baidu's release notes. Wedoany reports ERNIE 5.1's pre-training cost came in around 6 percent of the industry average. One tier down, the open spigot keeps flowing: Baidu open-sourced its Ernie-Image generation model on April 15, 2026, putting high-end rendering on consumer GPUs, per CnTechPost, and shipped an open multimodal reasoning model that VentureBeat notes Baidu claims beats GPT-5 and Gemini.
That word "claims" is doing real work. The benchmark numbers attached to these releases are vendor-published, and independent third-party audits of Chinese frontier models remain thin. Until named outside evaluators confirm the scores, treat them as marketing with error bars. Ecosystem costs are equally invisible on a price sheet: documentation gaps, Chinese-language support channels, and the engineering hours needed to self-host a 424-billion-parameter model.
What Does Chinese Law Mean for Your DeepSeek Data?
The legal dimension is fixed and does not depend on anyone's intentions. DeepSeek's own privacy policy states that user data, including prompts and device information, is stored on servers in the People's Republic of China. NPR and The Hill have detailed how China's National Intelligence Law of 2017 requires all Chinese organizations and citizens to support, assist, and cooperate with national intelligence work. That obligation applies regardless of what any company's privacy policy promises, and it sits alongside the data-localization and government-access provisions of China's Cybersecurity Law of 2017 and Data Security Law of 2021.
Regulators have acted on exactly this concern. Italy blocked the DeepSeek app in January 2025 after the company failed to answer GDPR questions, and Denmark banned it from government work devices, per State of Surveillance. The practical mitigation for developers is structural: open-weight Chinese models, such as the ERNIE 4.5 family, can be run on US-hosted infrastructure where prompts never touch Chinese servers. Using DeepSeek's $0.435 tokens through the company's own API is a categorically different risk decision, because no setting in the app removes the jurisdiction.
Will US Subscription Prices Follow China's Price War?
The pressure has already crossed the Pacific. On June 8, Google cut its entry-level Google AI Plus plan from $7.99 to $4.99 per month and doubled included storage to 400GB, per 9to5Google, a move TechCrunch called "a warning shot in the AI subscription price wars." When open-weight models hosted on commodity US clouds keep getting better, the floor under every Western rate card softens, and the consumer and indie-developer tiers feel it first.
The squeeze on OpenAI and Anthropic is real but asymmetric. Enterprise customers pay those companies for compliance, jurisdictional safety, support, and indemnification as much as for benchmark points, and none of those line items get cheaper because a Hangzhou lab cut token prices. The likelier near-term effect is exactly what Google just demonstrated: cuts at the entry level, where Chinese alternatives compete hardest, while flagship enterprise pricing holds.
The full decision has four parts; skipping any is how readers get burned. DeepSeek and its peers offer genuine, dramatic price advantages, roughly 29-fold on flagship output tokens. The benchmark claims behind those prices remain largely vendor-published and unverified by named independent auditors. The ecosystem carries hidden costs in documentation, support, and self-hosting engineering. And Chinese law, specifically the National Intelligence Law of 2017, gives Beijing on-demand access rights to data held by Chinese companies, a condition no contract can waive. Cheap tokens are real; so is everything they do not include.
Frequently Asked Questions
Is DeepSeek the most used AI in China?
Yes. DeepSeek's website recorded 541 million visits in May 2026, the most of any Chinese AI product and nearly three times second-place Nami AI Search's 198 million, according to traffic data reported by Pandaily. Globally it ranks fifth among AI products by web traffic.
How much cheaper is DeepSeek than OpenAI or Anthropic?
DeepSeek's V4-Pro API costs $0.435 per million input tokens and $0.87 per million output tokens, permanently set at one-quarter of its April launch price. OpenAI's GPT-5.2 lists at $1.75 and $14.00, and Anthropic's Claude Opus 4.8 at $5.00 and $25.00, making DeepSeek roughly 4 to 29 times cheaper depending on the comparison.
Is it safe to use the DeepSeek app in the US?
DeepSeek stores user data on servers in China, where the 2017 National Intelligence Law obligates companies to cooperate with state intelligence requests, and Italy and Denmark have restricted the app. Running open-weight Chinese models on US-hosted infrastructure avoids that jurisdiction; using DeepSeek's own app or API does not.
Did Baidu open-source its Ernie models?
Baidu open-sourced the ERNIE 4.5 family, ten models up to 424 billion parameters, under the Apache 2.0 license on June 30, 2025. Its current frontier models, ERNIE 5.0 and 5.1, remain proprietary and hosted-only, though Baidu did open-source its Ernie-Image model in April 2026.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
