AINews

GPT-4.5 : Technical Breakthroughs in OpenAI’s Latest Language Model

Following the powerful debut of Claude 3.7 Sonnet, OpenAI has officially released its latest language model, GPT-4.5, internally codenamed Orion. This model is positioned as the largest language model to date, focusing on enhancing unsupervised learning, multimodal processing, and emotional understanding capabilities. According to the OpenAI official announcement, GPT-4.5 aims to provide a more reliable and natural interactive experience, with breakthroughs in reducing hallucinations and enhancing human collaboration. This article will delve into GPT-4.5’s technical details, performance benchmarks, application scenarios, and potential impact on AI development.

Most importantly, Hong Kong users can also try it on POE as a Preview version, though it requires a subscription and each message consumes over 2,000 points.

Technical Core of GPT-4.5

GPT-4.5’s design incorporates multiple technical innovations, making it stand out in various application scenarios. Here are its main features and improvements:

Expansion of Unsupervised Learning

  • GPT-4.5 enhances its unsupervised learning capabilities to extract broader knowledge from unlabeled data, enabling it to handle more complex topics and adapt to diverse user needs.
  • Compared to previous generations, this technology improves the model’s “world understanding,” giving it more depth when generating contextually relevant responses.

Enhanced Emotional Understanding and Interaction

  • The model’s emotional quotient (EQ) has significantly improved, allowing it to more accurately identify user intentions and generate empathetic responses. For example, when faced with questions like “how to deal with stress,” GPT-4.5 can provide humanized advice.
  • Data shows its win rate in emotional understanding tasks compared to GPT-4o reaches 56.8% to 63.2%.

Reduced Hallucinations

  • Hallucinations (generating inaccurate or fictional information) have been a challenge for language models. GPT-4.5’s SimpleQA hallucination rate has decreased to 37.1%, lower than GPT-4o (44%) and other competing models.
  • This advancement improves its reliability in scenarios requiring high accuracy, such as academic research or professional writing.

Multimodal and Coding Capabilities

  • GPT-4.5 performs excellently in multimodal tasks (such as MMMLU combining images and text, scoring 85.1%) and coding tasks (such as SWE-Bench Verified, scoring 38.0%).
  • This makes it a powerful tool for developers and technical professionals, especially when generating code or processing complex data.

Performance Benchmarks: GPT-4.5 vs. Competitors

OpenAI provided multiple benchmark test results for GPT-4.5, compared with its previous model GPT-4o and the reasoning-specialized model o3-mini. Here are the key data:

TaskGPT-4.5GPT-4oo3-mini
GPQA (Science)71.4%53.6%79.7%
AIME ’24 (Mathematics)36.7%9.3%87.3%
MMMLU (Multimodal)85.1%81.5%81.1%
MMMU (Multimodal)74.4%69.1%Not provided
SWE-Lancer Diamond (Coding)32.6%23.3%10.8%
SWE-Bench Verified (Coding)38.0%30.7%61.0%

Analysis: GPT-4.5 performs exceptionally well in multimodal and coding tasks, particularly in MMMLU (85.1%) and SWE-Lancer Diamond (32.6%), surpassing GPT-4o. However, in reasoning-intensive tasks (such as AIME ’24 mathematics test, only 36.7%), o3-mini (87.3%) clearly has the advantage.

Positioning Difference: GPT-4.5 does not adopt the “think before answering” strategy of o1 or o3-mini, but instead focuses on versatility and quick response, suitable for multifunctional application scenarios.

Availability and Application Scenarios

Current Access Channels

ChatGPT Integration: GPT-4.5 is currently only available to ChatGPT Pro users for $200 per month. Features include file upload and search, but voice mode and video processing are not supported. OpenAI plans to expand to Plus/Team users next week, followed by Enterprise/Edu tiers.

API Functionality: Developers can access it through Chat Completions API, Assistants API, and Batch API, supporting function calls, structured output, and visual input (images). Its cost is higher than GPT-4o, with specific pricing still being adjusted.

Potential Applications

Content Generation: From technical articles to creative writing, GPT-4.5’s natural language generation capabilities can significantly improve efficiency.

Programming Support: Its performance in coding benchmarks makes it an ideal assistant for developers, such as generating code snippets or debugging complex projects.

Education and Research: Reduced hallucinations and powerful knowledge base make it useful for academic collaboration and knowledge organization.

Safety and Ethical Considerations

OpenAI emphasized GPT-4.5’s safety design, employing Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) technologies. According to its Preparedness Framework, the model has undergone rigorous testing to ensure outputs meet ethical and safety standards. This initiative responds to industry concerns about AI reliability and potential misuse.

Conclusion and Future Outlook

The release of GPT-4.5 demonstrates OpenAI’s continuous innovation in the field of language models. Its breakthroughs in unsupervised learning, multimodal processing, and emotional understanding make it a powerful tool for general AI applications. Although it doesn’t match o3-mini in reasoning tasks, its collaborative nature and reliability open new possibilities for professionals and developers. With the expansion of API capabilities and potential GPT-5 plans, GPT-4.5 may just be the starting point of OpenAI’s larger blueprint.

Reference: https://openai.com/index/introducing-gpt-4-5/