Alibaba has released Qwen4-Max, cutting the cost per task to RMB 11.16, down 58% from its predecessor. Tencent's Hy4 flash has pushed comparable task prices to as little as one-eighth of competing products, while DeepSeek has begun restricting free users because of insufficient compute. Together, these developments show that AI model competition is entering a phase defined by “lower prices for scale, with compute setting the ceiling.”

What This Is

This round of changes centers on model pricing and deployment costs. Qwen4-Max has 2.4T total parameters and supports native vision and million-character context, while Tencent's Hy4 flash activates only 12 billion parameters per call, making it an efficient model that invokes part of its parameters on demand. Both products focus on API pricing and the open sourcing of smaller models rather than simply expanding parameter counts.

Another development comes from Huawei Cloud. Its open-source CloudRobo-WM2.0 ranks first on the RoboTwin and LiberoPro robot manipulation benchmarks. Here, an “embodied world model” refers to a system that enables AI to understand its environment, predict the consequences of actions, and help robots perform tasks. Baidu, meanwhile, reported that its AI cloud revenue grew 94% year over year in the third quarter and revenue from token calls surged 1,159%, indicating that enterprise usage is increasing.

Industry View

What is encouraging is that falling prices and rising usage are occurring simultaneously, showing that model vendors are beginning to pass technical progress on to customers. Databricks' AIIndex, which covers more than 10,000 companies, also demonstrates that model adoption, spending, and the performance of coding agents can now be tracked continuously.

Compute supply, however, remains a hard constraint. DeepSeek's restrictions on long conversations and daily allowances for free users suggest that demand may be growing faster than data centers can expand. More alarming is SemiAnalysis' decision to cut its forecast for global AI accelerator capital expenditure by 28% for 2028 and 40% for 2030, which briefly sent Nvidia shares down more than 4%. This does not mean AI demand is disappearing; rather, the market is beginning to demand that enormous compute investments quickly generate returns. Anthropic's planned IPO will further test whether investors are willing to assign long-term value to model capabilities and commercialization prospects.

Impact on Regular People

For enterprise IT: Cheaper model calls will drive greater adoption in customer service, document processing, and internal knowledge bases. However, enterprises still need to assess peak-hour throttling, vendor lock-in, and data compliance.

For individual professionals: As the cost of agents—AI assistants that can autonomously complete multi-step tasks—falls, repetitive writing, programming, and analytical work will become easier to automate. More value will likely come from asking precise questions, checking results, and taking final responsibility.

For the consumer market: Low-cost models will accelerate the spread of free or inexpensive features, but tighter free allowances may also become a common business model. Consumers will need to reconsider the trade-offs among convenience, response speed, and paywalls.