At its Create 2025 conference in Wuhan last Friday, Baidu doubled down on its vision that the future of AI lies in creating smarter applications. Speaking to a packed room of developers, Baidu founder and CEO Robin Li said that as foundational models improve, they become a double-edged sword: while they unlock more possibilities for application development, they also risk making apps obsolete if the model itself renders them redundant. He pointed out that the key to building apps that stand the test of time is to “find the right scenario, pick the right model, and learn a little about fine-tuning.”
Lower cost than DeepSeek’s products
Baidu took direct aim at DeepSeek, one of its fiercest AI competitors, with the launch of two new large language models: Ernie 4.5 Turbo and X1 Turbo. Both offer multimodal reasoning capabilities and promise significant cost advantages. Ernie 4.5 Turbo costs about 40% of DeepSeek’s V3 model, while X1 Turbo undercuts DeepSeek’s R1 by offering prices at just 25% of its competitor’s rates.
“The future belongs to multimodal,” Li emphasized, arguing that pure-text models will shrink as multimodal models – those capable of handling text, images, and more – become the new standard.
Each million tokens on Ernie 4.5 Turbo now costs RMB 0.8 ($ 0.11) for input and RMB 3.2 ($ 0.44) for output, a massive discount compared to earlier models. X1 Turbo prices are similarly aggressive at RMB 1 ($ 0.14) for input and RMB 4 ($ 0.55) for output.
Baidu’s Manus competitor
Meanwhile, Baidu is racing into the agent space with Xinxiang, a new mobile app that mirrors Manus, the buzzy AI agent platform launched in March. Xinxiang allows users to input a task and have it executed via AI agents. Unlike Manus, which focuses on tool invocation via the MCP protocol, Xinxiang operates through what Baidu calls the Agent Use protocol – directly managing agents to fulfill tasks. Currently, Xinxiang supports around 200 types of tasks across work, education, and daily life. Baidu plans to scale this to over 100,000 tasks, with a PC version already in the works.
Pushing MCP as a new ecosystem
Baidu is also moving to build an open AI ecosystem. The company announced that its search engine will open up to third-party developers, encouraging them to build AI-powered apps within its search environment. It also added support for the MCP (Model Capability Protocol) standard, which enables seamless interaction between AI models, third-party tools, and databases – a move that could make Baidu’s platform more attractive to outside developers.
Competitors such as Alibaba Cloud and Tencent Cloud have already jumped on MCP to boost their own AI ecosystems, but Li believes the real competition will be won through applications, not just protocols. “There will be lots of models,” he said, “but it’s the apps that will rule.”
Baidu’s Wuhan conference attracted major tech partners including Intel and NVIDIA, along with emerging players such as Unitree Robotics and Leapmotor. The event also saw the launch of a new round of Baidu’s AI app competition, with the Chinese tech giant dangling RMB 70 million ($9.6 million) in prize money to spur innovation.