Baidu Slashes AI Costs with ERNIE 4.5 Turbo and X1 Turbo, Aiming at DeepSeek
Baidu Slashes AI Costs with ERNIE 4.5 Turbo and X1 Turbo, Aiming at DeepSeek
Baidu’s latest AI models—ERNIE 4.5 Turbo and ERNIE X1 Turbo—debuted at Create 2025 on April 25, offering dramatically lower pricing and competitive performance to challenge rivals like DeepSeek. ERNIE 4.5 Turbo operates at just $0.11 per million input tokens—an 80 percent reduction over its predecessor—and claims to outperform GPT-4o on multimodal benchmarks. Its companion, ERNIE X1 Turbo, brings advanced reasoning at $0.14 per million input tokens, reportedly 75 percent cheaper than DeepSeek R1. Alongside these models, Baidu introduced Xinxiang, a multi-agent system for 200+ tasks, and Huiboxing, a digital avatar platform. CEO Robin Li emphasized that text-only models are “shrinking” in relevance, as the company integrates these offerings across its cloud and consumer ecosystem, intensifying the AI cost-performance arms race in China and beyond .
Introducing ERNIE 4.5 Turbo and ERNIE X1 Turbo
Baidu’s new flagship models were unveiled at its annual Create 2025 developer conference in Wuhan on April 25, 2025:
- ERNIE 4.5 Turbo: A foundation model with enhanced multimodal capabilities—text, image, audio, and video—designed to rival GPT-4o on key benchmarks .
- ERNIE X1 Turbo: A reasoning-optimized model building on the March-launched X1 series, supporting complex tasks like document Q&A, code interpretation, and tool integration .
Key Features and Pricing
- Ultra-Low Input Costs
- ERNIE 4.5 Turbo: $0.11 per 1M input tokens (80 percent cheaper than its predecessor) .
- ERNIE X1 Turbo: $0.14 per 1M input tokens—approximately 25 percent of DeepSeek R1’s cost .
- Efficient Output Pricing
- X1 Turbo’s output rate: $0.55 per 1M tokens, maintaining cost-effectiveness for two-way interactions .
- Scalable Performance
- Both models leverage Baidu’s Kunlun P800 chip cluster, capable of training and serving hundred-billion-parameter models at scale .
Multimodal and Reasoning Capabilities
ERNIE 4.5 Turbo and X1 Turbo advance Baidu’s AI frontier by:
- Multimodal Understanding: Processing images, audio, and video alongside text, surpassing GPT-4o on internal benchmarks .
- Deep Reasoning: X1 Turbo integrates reflection and tool use for multi-step problem solving, outperforming DeepSeek R1 on reasoning benchmarks .
- Developer Tools: Integrated APIs for retrieval-augmented generation (RAG), code assistance, and domain-specific fine-tuning.
New AI Tools: Xinxiang & Huiboxing
Beyond model releases, Baidu showcased two supporting platforms:
- Xinxiang: A multi-agent orchestration system managing over 200 distinct tasks, from customer service to data analysis .
- Huiboxing: A digital avatar creation suite enabling interactive, personalized virtual assistants for marketing and CX use cases .
Market Impact: Challenging DeepSeek
Robin Li declared that “the market is shrinking” for text-only AI like DeepSeek R1, highlighting:
- Cost Pressure: Baidu’s models undercut leading competitors by 75–80 percent on core pricing .
- Performance Edge: Superior multimodal and reasoning benchmarks position ERNIE as a more versatile choice .
- Strategic Integration: Embedding ERNIE models into Baidu Cloud, Qianfan enterprise services, and consumer apps to drive adoption .
As China’s AI landscape heats up, Baidu’s aggressive pricing and feature push set a new bar for cost-effective, high-performance models—narrowing the gap with Western leaders and challenging domestic upstarts.
Comments
Post a Comment