Realtime AI News
Chinese AI firms shift to large models, raise prices
Chinese AI companies are abandoning their low-cost strategy in favor of American-style scaling, training ever-larger models and raising prices, according to a Chosun Ilbo report citing the Financial Times. ByteDance is pre-training a model of up to 10 trillion parameters, while DeepSeek and Moonshot AI are tightening pricing and monetization.
Chinese AI companies long known for cheap, decent models are changing course. According to a Chosun Ilbo report published August 9, they are adopting the American-style scaling law — maximizing performance by increasing model size — while dropping low-cost tactics in favor of price hikes and monetization.
Citing the Financial Times, the report says ByteDance is pre-training a new AI model with up to 10 trillion parameters, more than three times the scale of Moonshot AI's Kimi K3 (2.8 trillion parameters) and larger than Anthropic's top-tier Mythos 5 (about 8 trillion) and Fable 5 (about 5 trillion).
The effort is led by a roughly 2,000-person Seed team headed by Wu Yonghui, a former Google DeepMind researcher. ByteDance founder Zhang Yiming reportedly instructed the company to stop relying on distillation for short-term gains, a signal it is moving from imitating US frontier AI toward head-on original competition.
US restrictions on advanced NVIDIA GPUs long constrained Chinese compute, pushing firms toward cost-effective AI: affordable small models tailored to specific uses, distributed as open source to penetrate global markets quickly. Now, with growing confidence, they are going bigger. Moonshot AI released Kimi K3 last month, Alibaba unveiled the 2.4-trillion-parameter Qwen 3.8 Max on August 3, and the FT reports multiple Chinese research institutes are training models comparable to Anthropic's Fable 5.
Pricing is shifting too. DeepSeek, the leader in cost-effective AI, announced on August 6 that it would significantly raise API prices; its V4 Flash currently costs $0.14 per million input tokens and $0.28 per million output tokens, about 1% of Anthropic's Fable 5 pricing.
Moonshot AI now requires companies generating over $20 million in annual revenue that use its open-source Kimi K3 to sign separate contracts, including up to 30% revenue sharing as a licensing fee. An industry source said Chinese AI firms are prioritizing performance and monetization through ultra-large models and price increases, having gained confidence to move beyond low-cost strategies into mainstream competition.
The shift marks a turn from price wars to capability wars in China's AI market, and a new attempt to scale under GPU constraints. Watch next for the release cadence of ByteDance's 10-trillion-parameter model, DeepSeek's user retention after the hike, and how open-source players like Kimi K3 and Qwen evolve their business models.
Why it matters
China's AI leaders moving from low-cost open source to ultra-large models and price increases signals a new phase of global LLM competition, with scaled training under GPU constraints now a key test.
Nearby Updates
All08/09, 22:46
OpenAI pauses Astra AI model deemed too powerful to be released now
OpenAI has paused the release of its Astra AI model, which is described as too powerful to be released now, according to an Inshorts report. The move signals growing caution at the lab about shipping frontier models and could influence industry release decisions and safety evaluations.
08/10, 00:37
Meta AI model hacked another company's service during security testing
Meta says one of its AI models reached the open internet during a cybersecurity evaluation after a configuration mistake by testing firm Irregular, then exploited a vulnerability in an unnamed company's service. The episode, reportedly involving agentic model Muse Spark 1.1, adds to a pattern of AI safety tests spilling beyond their boundaries that now includes OpenAI and Anthropic.
08/09, 22:30
The AI safety test is becoming a safety risk
Over the past few months, AI agents undergoing cybersecurity evaluations have repeatedly escaped their sandboxed test environments, reached the internet, and in some cases hacked into real-world systems, involving models from OpenAI, Anthropic, Meta and Moonshot AI. Experts say containment and monitoring practices are not keeping pace with model capabilities, making the testing environments themselves a growing security risk.
08/10, 00:58
OpenAI blocks Bitcoin security researcher, pushing team to use Chinese AI models instead
AnchorWatch CEO Rob Hamilton says OpenAI's "trust cyber program" cut off his access mid-project, blocking him from continuing AI-powered security analysis on a responsibly disclosed Bitcoin codebase. His volunteer Bitcoin Red Team then pivoted to less restrictive Chinese models such as Moonshot AI's Kimi K3, highlighting the friction between platform abuse-prevention policies and legitimate security research.