Baidu's ERNIE 4.5 & X1: Features, Access, DeepSeek Comparison

This isn’t Baidu’s first step into AI. The company, originally founded in 2000 as China’s leading search engine (often called the “Google of China”), began seriously investing in artificial intelligence several years ago. Baidu has been developing ERNIE (Enhanced Representation through Knowledge Integration) since 2019, and they released the ERNIE Bot in 2023.

However, Baidu’s dominance in the Asian AI market has recently come under pressure from Alibaba’s Qwen models and innovative new competitors like DeepSeek.

With ERNIE 4.5, Baidu directly competes with DeepSeek’s V3, Qwen 2.5 Max, or OpenAI’s GPT-4o.

What Is Baidu’s ERNIE X1?

ERNIE X1 is a specialized reasoning model designed for advanced tasks such as mathematics and complex coding. Similar to DeepSeek-R1 or OpenAI’s o1, this type of model explicitly shows its thought process to the user, making it easier to understand how it arrives at a particular answer or solution.

Companies continue investing heavily in reasoning-focused models like ERNIE X1 because these models directly target tasks that bring clear business value. According to recent data from Anthropic, reasoning and coding tasks are the top enterprise use cases, making up a significant portion of AI adoption.

Source: The Anthropic Economic Index

Despite AI’s rapid development, enterprise adoption remains low globally, and models that excel in reasoning-intensive domains (math, coding, advanced analytics) are more likely to deliver tangible business value, driving revenue and accelerating AI adoption in the enterprise market.

Baidu’s main selling point for ERNIE X1 is its aggressive pricing strategy. Here’s a quick comparison to DeepSeek-R1 (pricing information is accurate as of March 17, 2025):

At standard rates (and excluding content caching), Baidu’s claim—that ERNIE X1 matches DeepSeek-R1 performance at half the price—holds true, particularly regarding output token pricing.

However, under DeepSeek’s discounted rates (which apply during 16:30-00:30 UTC each day), the reverse claim applies: ERNIE X1 becomes twice as expensive.

Despite attractive pricing, Baidu has yet to provide detailed benchmarks confirming ERNIE X1’s performance relative to DeepSeek-R1, which leaves some uncertainty about its real-world competitiveness.

While we wait for benchmarks on ERNIE X1, let’s find more details on the performance of ERNIE 4.5.

ERNIE 4.5 Benchmarks

Baidu released benchmarks comparing ERNIE 4.5 to leading multimodal and text-based models, including OpenAI’s GPT-4o, GPT-4.5, and DeepSeek’s V3. Let’s walk through the key results, starting with multimodal capabilities.

Multimodal benchmarks

In multimodal benchmarks, ERNIE 4.5 outperformed GPT-4o, achieving an average score of 77.77, compared to GPT-4o’s 73.92—a lead of 3.85 points.

Source: Baidu

Specifically, ERNIE 4.5 surpassed GPT-4o in six out of seven evaluated benchmarks:

• CCBench: Evaluates common-sense reasoning across text and images. ERNIE 4.5 scored approximately 81, slightly outperforming GPT-4o’s ~79.
• OCRBench: Assesses optical character recognition capabilities, focusing on text extraction from images. ERNIE 4.5 achieved around 88, surpassing GPT-4o’s ~81.
• ChartQA: Tests understanding of data presented in charts. ERNIE 4.5 scored ~82, marginally ahead of GPT-4o’s ~81.
• MMMU: Measures multimodal reasoning across various topics. Here, GPT-4o led with ~70, while ERNIE 4.5 scored ~64, indicating an area for improvement.
• MathVista: Evaluates mathematical reasoning in visual contexts. ERNIE 4.5 scored ~69, outperforming GPT-4o’s ~61.
• DocVQA: Assesses the ability to answer questions based on document visuals. ERNIE 4.5 excelled with a score of ~91, compared to GPT-4o’s ~85.
• MVBench: Focuses on temporal understanding in dynamic video tasks, requiring reasoning over sequences of frames. ERNIE 4.5 scored ~72, significantly outperforming GPT-4o’s ~63.

Text-only benchmarks

On text-only tasks, ERNIE 4.5 achieved an average score of 79.6, slightly ahead of GPT-4.5’s average of 79.14, and also surpassing DeepSeek-V3 (~77).

Source: Baidu

These results suggest that ERNIE 4.5 is particularly competitive in multimodal document understanding, mathematics, and Chinese language benchmarks. However, its weaker performance on specific coding benchmarks, such as LiveCodeBench, highlights potential limitations in specialized coding tasks.

Let’s quickly break down a few key benchmarks:

• MMLU-Pro: Assesses multitask learning across various disciplines. ERNIE 4.5 scored ~78, slightly below GPT-4.5’s ~79.
• GPQA: Evaluates general-purpose question answering. ERNIE 4.5 achieved ~57, trailing GPT-4.5’s ~61.
• C-Eval: Measures Chinese general knowledge and reasoning. ERNIE 4.5 led with ~88, outperforming GPT-4.5’s ~80.
• CMMLU: Tests Chinese-language multitask understanding. ERNIE 4.5 scored ~88, ahead of GPT-4.5’s ~80.
• Math-500: Measures a model's ability to solve challenging high-school-level mathematical problems. ERNIE 4.5 (~82) is behind the leader DeepSeek-V3 (~88) and GPT-4.5 (~84).
• CMath: Evaluates math problem-solving in Chinese. ERNIE 4.5 led with ~95, surpassing DeepSeek-V3’s ~85.
• LiveCodeBench: Measures real-time coding abilities. ERNIE 4.5 scored ~35, below GPT-4.5’s ~45, indicating room for improvement in coding tasks.

How to Access ERNIE 4.5 and ERNIE X1

You can directly try Baidu’s ERNIE 4.5 and X1 through the official chatbot application available on their website: yiyan.baidu.com.

However, based on my personal experience, there are a few usability issues. The interface is currently available primarily in Chinese, making it challenging for non-Chinese speakers—while you can use the browser’s automatic translation feature, the translation is not smooth and the UX suffers.

Source: Baidu

Additionally, Google or GitHub login is unavailable, a problem I haven’t encountered with competitors like DeepSeek or Qwen. I still tried to create an account, but I wasn’t allowed to complete the registration form with my European phone number.

How to access ERNIE via API

For developers interested in integrating ERNIE 4.5, API access is already available via Baidu’s official platform, Qianfan. Pricing starts at $0.55 per 1 million input tokens and $2.20 per 1 million output tokens.

As of March 17, 2024, ERNIE X1 is not yet available through the API, but Baidu mentioned that it will be soon.

Moreover, Baidu has announced plans to open-source ERNIE 4.5 starting June 30, 2025. This step could improve adoption internationally by making the model more accessible and adaptable.

Baidu also intends to integrate ERNIE 4.5 and X1 into various products across its ecosystem—including Baidu Search and the Wenxiaoyan App—though timelines and details remain to be confirmed.

ERNIE’s Impact on the AI Market

The release of Baidu’s ERNIE 4.5 and ERNIE X1 continues a notable pattern I’ve seen emerging in China’s AI industry—a strategy of prioritizing disruption over immediate usability.

Unlike companies such as Google, OpenAI, or Mistral, which typically spend extensive periods (often 8–12 months or more) ensuring their products meet stringent stability, security, privacy, and safety standards, Chinese AI companies seem to favor rapidly deploying products that are compelling yet still in an early, unpolished state.

Just like ManusAI and DeepSeek, Baidu’s ERNIE models follow a similar trend. They enter the market aggressively, aiming primarily to disrupt pricing structures and competitive assumptions. The ERNIE X1 model, in particular, explicitly positions itself as comparable in performance to DeepSeek-R1 at half the cost—even though Baidu hasn’t yet publicly backed these claims with detailed benchmarks.

And, while ERNIE 4.5 shows genuinely competitive benchmark results, international users like myself still encounter barriers, notably a challenging sign-up process and language accessibility issues.

Overall, the ERNIE 4.5 and X1 releases reinforce the idea that Chinese AI companies are comfortable sacrificing some stability and user experience in the short term, as long as they’re able to quickly push the market forward and provoke industry-wide conversations.

Whether this approach is sustainable remains to be seen, but there’s no denying it’s effective at forcing the global AI market to move faster and think differently about what it means to compete in the current market.

Conclusion

Baidu’s release of ERNIE 4.5 and ERNIE X1 represents yet another ambitious step in China’s competitive AI landscape. ERNIE 4.5 demonstrates clear strengths, particularly in multimodal understanding, mathematics, and Chinese-language tasks, positioning it strongly against global rivals like GPT-4o and DeepSeek V3.

On the other hand, ERNIE X1’s promise of delivering advanced reasoning capabilities at half the cost of DeepSeek-R1 is intriguing—but without benchmarks, it remains speculative.

As these models mature, it’ll be interesting to see whether Baidu and similar Chinese companies adjust their approach to strike a better balance between disruption and stability.