Baidu Escalates China’s AI Price War with Faster, Cheaper ERNIE Turbo Models

Baidu has intensified the ongoing price war in China’s artificial intelligence sector, unveiling significantly cheaper and faster versions of its large language models during its Baidu Create 2025 developer conference in Wuhan today. The introduction of ERNIE X1 Turbo and ERNIE 4.5 Turbo underscores Baidu’s strategy to gain ground against domestic competitors like DeepSeek by lowering the cost barrier for AI development. Reflecting market reception, Baidu shares saw a notable rise following the news, climbing as much as 5.4 percent in Hong Kong trading.

The announcement comes just over a month after Baidu launched the original ERNIE 4.5 and X1 models in March, which were already positioned aggressively on price. The original ERNIE 4.5, noted at the time for strong multimodal performance surpassing competitors like GPT-4o in some tests, had an API access fee of around $0.55 per million input tokens and $2.20 per million output tokens via Baidu Qianfan.

Now, ERNIE 4.5 Turbo, an upgrade to this foundation multimodal model capable of processing text and images, boasts faster processing and an 80% price reduction from that previous baseline. The new Turbo version is priced at RMB 0.8 per million input tokens and RMB 3.2 per million output tokens (approximately $0.11/$0.44 USD), making it about 40% of the cost of DeepSeek V3, according to Baidu.

Improvements cited for 4.5 Turbo include better multimodal reasoning, logical reasoning, coding output, and fewer instances of hallucination. Further signaling a push for wider use, Baidu plans to open-source the ERNIE 4.5 Turbo model code in June.

Alongside this, Baidu introduced ERNIE X1 Turbo, enhancing its deep-thinking reasoning model. A reasoning model focuses on solving complex, multi-step problems requiring logical analysis. Baidu claims X1 Turbo can outperform DeepSeek R1 and the latest version of DeepSeek V3, featuring what the company calls a more advanced thinking chain and stronger deep thinking ability.

Its pricing undercuts DeepSeek R1 significantly, set at RMB 1 per million input tokens and RMB 4 per million output tokens (approximately $0.14/$0.55 USD), representing about 25% of R1’s cost. Baidu highlighted its strengths in complex text generation, logic puzzles, math, literary creation, and image understanding.

Strategic Push for Accessibility

This aggressive pricing follows Baidu’s move earlier this year to make its flagship AI chatbot, Ernie Bot, completely free, scrapping a previous ~$8 monthly fee. That decision was seen as a direct response to market pressure from rivals like ByteDance’s Doubao, Alibaba’s Qwen, and DeepSeek, amidst reports suggesting Baidu needed to boost user adoption.

According to March 2025 data, Ernie Bot ranked eighth among Chinese chatbots by monthly active users, highlighting the competitive field Baidu operates in within a market estimated to have over 130 large language models.

CEO Robin Li reiterated the focus on affordability at the conference, stating, “Why do we keep lowering the price of large language models? Because it’s a big obstacle for developers to make AI applications because of the high cost, they can’t afford it.”

He also emphasized the growing importance of multimodal capabilities, stating that multimodality is the future standard driven by user demand, predicting continued market growth while demand for pure text models subsides, as per AASTOCKS.

Expanding the AI Ecosystem

Beyond the model upgrades, Baidu used the Create event to showcase broader ambitions. The company debuted “Xinxiang,” an AI agent development platform designed to automate tasks and facilitate the building of complex AI applications.

AI agents are software programs designed to perceive their environment and take actions to achieve specific goals autonomously. Baidu also announced initiatives aimed at nurturing the developer community, including the third “Wenxin Cup” entrepreneurship competition with substantial investment prizes and a goal to cultivate 10 million AI professionals over the next five years.

Highlighting its infrastructure development, Baidu unveiled a domestically developed AI training cluster utilizing 30,000 cards. The ERNIE series, built upon Baidu’s internal architecture incorporating features like retrieval-augmented generation (RAG) for accessing real-time data, underpins many of the company’s core products like Baidu Search and Maps. RAG allows models to pull in information from external, up-to-date knowledge sources to improve the accuracy and relevance of their responses.

The competitive landscape remains dynamic. DeepSeek is preparing its R2 reasoning model, and global players like OpenAI, Google and Anthropic continue setting benchmarks. Baidu itself is already working on ERNIE 5, expected in the latter half of 2025, which is anticipated to further enhance capabilities like real-time video processing and logical inference.

While Baidu’s aggressive pricing and performance upgrades aim to capture more users and developers, the efficiency and real-world utility of these Turbo models compared to rapidly evolving competitors will be closely watched. Potential increases in power consumption associated with enhanced capabilities also remain an area for observation in real-world deployments.