Image source: Generated by Unbounded AI

On February 28th, OpenAI today announced the launch of its latest flagship product GPT-4.5 language model.

This AI system, known as "the richest knowledge reserves to date", will be the first to open research previews to Pro users and developers. The official stated that this model is not a cutting-edge architecture version and may not perform as well as cutting-edge models such as o1 or o3-mini in some scenarios.

From now on, ChatGPT Pro users can select GPT-4.5 in the model selectors on the web, mobile and desktop. The model will begin rolling out to Plus and Team users next week, and then to corporate and educational users in the following week.

GPT-4.5 can access the latest information, support search, file and image uploads, and can write and code editing through canvas. It currently does not support multimodal functions in ChatGPT, such as voice mode, video and screen sharing.

In addition, OpenAI also previews GPT-4.5 to developers at all paid user levels in the Chat Completions API, Assistants API, and Batch API, supporting key features such as function calls, structured output, streaming and system messages. It also supports visual capabilities through image input.

OpenAI pointed out that GPT-4.5 is a very large and computationally intensive model that costs more than GPT-4o and is not a replacement for the latter.

GPT-4.5 has made significant breakthroughs in pre-training and post-training. By expanding the scale of unsupervised learning, GPT-4.5 improves the ability to identify patterns, build connections, and generate creative insights without direct reasoning.

Early tests showed that users felt more natural when interacting with GPT-4.5. Its broader knowledge base, improved understanding of user intentions, and a higher “emotional intelligence” (EQ) make it outstanding in tasks such as improving writing, programming, and solving practical problems. Meanwhile, OpenAI expects GPT-4.5 to have less fictional information, even slightly lower than its o1 model.

In SimpleQA (Simple Q&A) test, GPT-4.5 performed excellently in both accuracy and fictitious information rate

According to information disclosed by Ultraman Sam, OpenAI plans to launch GPT-4.5 before the end of February and GPT-5 before the end of May.

OpenAI CEO Sam Altman once called GPT-5 "a system that integrates many of our technologies" and will include new o3 reasoning that OpenAI showed at the Christmas 12th event last DecemberModel.

Previously, OpenAI released the o3-mini, but the full o3 model will be launched only as part of the upcoming GPT-5 system. This is consistent with OpenAI's goal to integrate its large language models to ultimately create a more powerful model that is expected to be seen as general artificial intelligence (AGI).

Unsupervised learning improves model accuracy and intuition

OpenAI improves AI capabilities by extending two complementary paradigms: unsupervised learning and reasoning, which represent two different dimensions of intelligence.

The extension of reasoning ability teaches models to think and generate chains of thought before responding, allowing them to solve complex scientific, technical, engineering and mathematics (STEM) or logic problems. Models such as OpenAI o1 and OpenAI o3min are representative of this paradigm.

Relatively speaking, unsupervised learning improves the accuracy and intuition of world models.

When asked about GPT-4.5 "what the first language is," it gave the answer: "We are not yet fully aware of what the first language is and cites perspectives for users' reference.

GPT-4.5 achieves the expansion of unsupervised learning by extending the scale of computing and data, as well as architectural and optimization innovations. As a result, the model has a broader knowledge and a deeper understanding of the world, thus reducing fictional information and improving reliability on a wide range of topics.

Conversation is more in line with the needs of human collaboration

As OpenAI researchers scale up models and solve more complex problems, it is becoming increasingly important to teach them to better understand human needs and intentions.

For GPT-4.5, researchers have developed new scalable technologies that enable the use of data from smaller models to train larger and more powerful models. These technologies improve the manipulation of GPT-4.5, understanding of nuances, and the ability to natural dialogue.

In comparison with human testers, GPT-4.5 has shown a high winning rate in creative intelligence, professional queries, and daily queries

Combining a deep understanding of the world with improved collaboration capabilities, GPT-4.5 can naturally integrate ideas in warm and intuitive conversations that are more in line with the needs of human collaboration.

class="ne-empty-p">

GPT-4.5's answer is more like a normal person

OpenAI demonstrates the results of GPT-4.5 and GPT-4o dealing with three identical problems:

Problem 1: Help in difficult times

The user said, "I failed the exam and it's very sad now." The answer to GPT-4.5 has better emotional substitution, and compared to the 4o mechanical solution, the "AI flavor" is much less.

‌Question 2: Space Exploration

Users‌ asked, “What do you think of space exploration?” GPT-4.5’s answer is more like a normal person, outputting his own opinions and opinions, while GPT-4o’s answer quotes classics and is too pedantic.

Inference ability will be the core capability of future models

GPT-4.5 does not perform pre-inference in response, which makes its advantages completely different from OpenAI's inference models such as o1. Compared with OpenAI's o1 and o3-mini, GPT-4.5 is a more general and smarter model.

OpenAI believes that reasoning capabilities will be the core capabilities of future models, and the two extension methods of pre-training and inference will complement each other.

As models like GPT-4.5 become smarter and more knowledgeable through pre-training, they will be a stronger foundation for reasoning and tools to use agents.

The improvement of model capabilities brings about security improvements

Every improvement of model capabilities is also an opportunity to make it safer. GPT-4.5 adopts new supervised training techniques, combining traditional supervised fine-tuning (SFT) and human feedback reinforcement learning (RLHF) methods, similar to those used by GPT-4o. OpenAI hopes that these efforts will provide the basis for more powerful model alignment in the future.

To test OpenAI improvements, the researchers conducted a series of security tests based on their preparation framework before deployment. It was found that extending the GPT paradigm helped to improve the capabilities of GPT-4.5 in various assessments. (Tencent Technology's special compiled by Wuji)