ERNIE 4.5 + X1: Most POWERFUL and CHEAPEST LLM Beats GPT-4.5, R1, & Sonnet 3.7! (Fully Tested)

Key Concepts:

• Ernie 4.5: A natively multimodal model by Baidu (BU) that excels in understanding, generation, reasoning, memory, hallucination prevention, logical reasoning, and coding.
• Ernie X1: Baidu's first reasoning AI model, designed for problem-solving, logical deduction, and complex reasoning tasks.
• Multimodality: The ability of a model to process and understand different types of data, such as text, images, audio, and video.
• Cost-efficiency: The models are priced significantly lower than competitors like GPT-4.5 and DeepSeek R1.
• Open Source: Baidu plans to make Ernie X1 open source later in the year (from June 30th), allowing for community contributions and wider accessibility.
• Reasoning AI: AI models designed to perform complex reasoning tasks, solve problems, and make logical deductions.
• Hallucination Prevention: The ability of a model to avoid generating false or nonsensical information.

Ernie 4.5: Multimodal Performance and Capabilities

• Overview: Ernie 4.5 is a multimodal model that demonstrates improvements in understanding, generation, reasoning, and memory. It also features enhanced hallucination prevention, logical reasoning, and coding abilities.
• Coding Test 1: Personal Finance Tracker App: The model was tasked with building a responsive web app to track monthly income and expenses. It successfully generated functional code, creating an app with expense visualization capabilities.
  • The app allowed users to input expenses, select categories, and visualize their financial data.
• Coding Test 2: Conway's Game of Life: Ernie 4.5 was asked to create Conway's Game of Life in Python, testing its ability to implement grid-based simulations and efficient state updates.
  • The model successfully generated the game, demonstrating its proficiency in simulation implementation.
• Coding Test 3: SVG Butterfly: The model was challenged to create an SVG representation of a butterfly with symmetrical wings and simple styling.
  • Ernie 4.5 generated a good SVG clip art of a butterfly with symmetrical designs on the wings and antenna. This is a difficult task, with only a few models like Claude 3.7 Sonnet and Gemini 2.5 Pro succeeding.
• Multimodal Example: Image Explanation: The model was given an image and asked to explain it, showcasing its vision reasoning capabilities.
  • The model was able to analyze the image and provide a detailed explanation of its contents.
• Debugging Test: The model was requested to identify and fix logical errors in a Python function.
  • Ernie 4.5 successfully identified and corrected the errors, providing the correct Python code and explaining the fixes.

Ernie X1: Reasoning AI Model

• Overview: Ernie X1 is Baidu's reasoning AI model, designed for problem-solving, logical deduction, and complex reasoning tasks. It matches DeepSeek's R1 performance at half the price.
• Word Problem Test: Train Meeting Point: The model was given a word problem involving two trains traveling towards each other and asked to determine when and where they would meet.
  • The model correctly calculated that the trains would meet at 11:59 a.m., approximately 261.3 km from city A.
  • The model showcased its reasoning process, demonstrating how it arrived at the correct answer.
• Geometric Equation Test: Triangular Field Division: The model was asked to find the length of a dividing line that would split a triangular field into two equal areas.
  • The model correctly calculated the length of the dividing line as 12 meters, using appropriate formulas.
• Integer Solution Test: Library Purchase: The model was tasked with finding different combinations of workbooks, educational app licenses, and science kits that would total exactly $250, with at least one of each item purchased.
  • The model correctly determined that there was no combination of items that would equal $250, even after applying brute force methods.
• Logical Deductive Reasoning Test: Truth Tellers and Liars: The model was presented with a scenario involving truth tellers and liars and asked to identify who was who based on their statements.
  • The model correctly identified A as the liar, and B and C as truth tellers, demonstrating its logical reasoning abilities.

Pricing and Accessibility

• Ernie 4.5 Pricing: $0.55 per 1 million input tokens and $2.20 per 1 million output tokens.
• Ernie X1 Pricing: $0.28 per 1 million input tokens and $1.10 per 1 million output tokens (50% of DeepSeek R1's price).
• Accessibility: Both models are freely accessible to all Erniebot users, and creating an account is free.
• Open Source: Ernie X1 will be made open source later in the year (from June 30th).

Conclusion

Baidu's Ernie 4.5 and Ernie X1 models represent significant advancements in AI, offering high performance at a lower cost compared to competitors. Ernie 4.5 excels in multimodal tasks, while Ernie X1 demonstrates strong reasoning capabilities. The cost-efficiency and planned open-source release of Ernie X1 are expected to drive broader adoption and innovation in the AI field. Robert Lee, the CEO of Baidu, emphasized that cost reduction and continuous innovation are crucial for AI scalability and real-world impact.