GuShiio.com 
(Photo source: Photo taken by Zhijia Lin, editor of GuShiio.comAGI)
At 4 o’clock this morning, GPT-4.5 suddenly went online.
On February 27, GuShiio.comAGI learned thatOpenAI Company in the United States today launched the GPT-4.5 model (codenamed Orion), which is the company’s largest and best AI base model with data scale in history.
GPT-4.5 was not an inference model from the beginning. OpenAI said that GPT-4.5 is a step forward in extending pre-and post-training. By extending unsupervised learning, GPT-4.5 improves the ability to recognize patterns, make connections, and generate creative insights without reasoning. In addition, GPT-4.5 is more than 10 times more efficient than GPT-4. At the price level, the GPT-4.5 API inputs US$75 per million tokens and outputs US$150. This is 30 times higher than the $2.50 for GPT-4o and 25 times higher than competitor Claude 3.7 Sonnet.
Although OpenAI CEO Sam Altman was not present at the conference, he tweeted that GPT‑4.5 made him feel like he was talking to a thoughtful person for the first time and could get really good advice from the model. But the bad news is that this is a huge and expensive model and I really want to launch both Plus and Pro versions, but we don’t have enough GPU computing cards. Next week we will add tens of thousands of GPU cards.& rdquo;
“(I) take care of my children in the hospital. The team has succeeded!& rdquo; Altman said.
The strongest model released in 6 years: more expensive, higher emotional intelligence, less hallucinations
It is reported thatIt has been six years since GPT-1 to the upcoming GPT-4.5.
In 8,In June, OpenAI released GPT-1, which is OpenAI’s first large-scale pre-trained language model based on the Transformer architecture; in 2019, OpenAI released GPT-2, which expanded the model scale 10 times and has 150 million parameters. It shows strong capabilities in generating text, but due to potential abuse risks, OpenAI uses it in internal testing.
In 0,In May, OpenAI launched GPT-3, which has 175 billion parameters and performs amazingly in natural language processing tasks. It can complete various tasks such as generating text, answering questions, and translating. By 2022, GPT-3.5 will be released, and OpenAI uses manual annotation of data and reinforcement learning to improve model performance.On November 30 of the same year, ChatGPT, an AI chat robot product based on GPT-3.5, was released and became popular around the world.
On March 14, 2023,OpenAI released GPT-4, which has stronger language understanding capabilities and can process image content. It is open to Plus users with a monthly subscription fee of US$20. Later in November, OpenAI announced the upgrade of GPT-4 to GPT-4 Turbo at the first developer conference.
By 4,In May, OpenAI launched a free-to-use multimodal model GPT-4o, and launched GPT-4o mini on July 18; on September 12, OpenAI officially released a preview version of the o1 model and also released o1-mini. In addition, on December 5, OpenAI released the official version of the OpenAI o1 model, and later announced the o3-mini series, which surpassed the o1 model in performance and cost performance.
However, under the influence of the open source AI model DeepSeek V3/R1 and Musk’s bidding action, everything changed on February 13 this year. OpenAI finally stopped squeezing toothpaste, and the entire product line was fully accelerated. The GPT-5 model will be released as soon as this year.
Altman admitted in a tweet that OpenAI has realized that its model and product supply have become very complex and needs to simplify product supply. ldquo; We hate model selection as much as you do and want to return to magical unified intelligence. Our primary goal is to unify the o-series models and the GPT series models by creating a system that can use all of our tools, knows when it takes long thinking, and can often be used for a very wide range of tasks. rdquo; Altman said.
Altman said that OpenAI will soon (within a few weeks/months) release GPT-4.5 codenamed Orion, which is the last non-thought chain model, and will integrate GPT and the o series, and will soon (within a few months) launch GPT-5 with multiple new features.
Altman emphasized that the previously announced reasoning model O3 will not be released as a stand-alone model. Most importantly, the free version of ChatGPT can use GPT-5 basic classes for conversations without restrictions under standard smart settings, but it will prevent abuse, while Plus/Pro paying users will use GPT-5 with a higher level of intelligence.Clearly, GPT-5 will also become the company’s first world model.
Today, OpenAI is the first to release GPT-4.5, the company’s largest, most expensive model with higher EQ and fewer illusions in six years.
OpenAI said that GPT-4.5 has made progress in extending pre-training and post-training, improving capabilities such as pattern recognition by extending unsupervised learning.
In terms of ability improvement,Early testing showed that GPT-4.5 interacts more naturally, has a broader knowledge base, can better understand user intentions, has higher emotional intelligence, reduces hallucinations, and performs well in tasks such as writing, programming, and solving practical problems. In the SimpleQA (Assessing Model’s Factual Answer Ability) dataset test,The accuracy rate of GPT-4.5 reaches 62.5%, which is higher than that of GPT-4o, o3 mini series, etc.; the illusion rate is as low as 37.1%, which is far superior to GPT-4o, etc.
same time,GPT-4.5 has unsupervised learning extensions. By expanding calculations, data, architecture and optimization innovations, it improves the accuracy and intuition of the world model, has broader knowledge and a deeper understanding of the world, and uses small model data to train large models, improves GPT-4.5 ‘s controllability, understanding of nuances and natural dialogue capabilities. Moreover, the training uses new supervision technologies, combines traditional methods, and conducts security testing before deployment. Relevant evaluation results will be released in the system.
In a comparative evaluation with human testers, GPT-4.5 has a higher win rate than 4o in terms of creative intelligence, professional query and daily query, showing stronger aesthetic intuition and creativity, and can reach 57% in daily query, professional query up to 63.2%. In addition, although GPT-4.5 does not have in-depth thinking, reasoning will become the core ability of the model in the future, so GPT-4.5 uses two extended methods of pre-training and reasoning to complement each other.
At the usage level,ChatGPT Pro users can choose it in the Model Selector from now on, will be available to Plus and Team users from next week, and will be available to Enterprise and Edu users next week. This version supports searching to obtain the latest information, uploading files and images, and using canvas to process writing and code, but does not support multimodal functions such as voice, video and screen sharing; At the API level, previews are available to all paying developers in the Chat Completions API, Assistants API, and Batch API. Key functions such as function calls and image input visual functions are supported, and are suitable for application scenarios such as writing assistance. However, due to the large, computation-intensive model and high cost, officials are evaluating whether to provide it in the API for a long time.
Box AI CEO Aaron Levie said he will launch GPT-4.5 version to customers in Box AI Studio later today. Through early testing, compared with GPT-4o, GPT-4.5 has increased the accuracy of correctly extracted fields by 19 percentage points, highlighting its improved ability to process subtle contract data. It is seen that GPT-4.5 has achieved strong results in processing complex enterprise data, which will unlock more use cases in the enterprise.
Scott Wu, co-founder and CEO of Cognition, shared his experience of using GPT-4.5 and said it was great. In their agent coding benchmark, GPT-4.5 achieved significant improvements compared to o1 and 4o. An interesting data point was also found: Although GPT-4.5 and Claude 3.7 Sonnet scored similarly on the overall benchmark, they found that GPT-4.5 peaked more on tasks involving architecture and cross-system interactions, while Claude 3.7 Sonnet peaked more on raw coding and code editing.
OpenAI said that GPT-4.5 is at the forefront of unsupervised learning and cannot yet completely replace GPT-4o.
OpenAI will collide with kimi and DeepSeek at the same time”
In fact,Before November 30, 2022, OpenAI’s website traffic will be almost zero. But in the following two months, OpenAI was hit by more than 100 million visitors, and everyone rushed to experience ChatGPT. Since then, everyone’s lives have been different, especially the company’s CEO Altman, who has become an AI technology evangelist and industry guide.
Today, OpenAI is valued at more than US$157 billion (approximately RMB 1.1 trillion)
Recently, Altman publicly stated that OpenAI is considering pricing based on usage. As for when AGI can be implemented, he said that when an AI system can do what a very skilled person can do in important work, it can be called AGI.
Interestingly, early this morning, kimi on the dark side of the moon crashed again and quietly announced the latest Kimi-K1.6-IOI-high model, which ranked first on the LiveCodeBench benchmark list, surpassing the GPT and Claude series models.
At the same time, DeepSeek Open Source Week continues to attract attention, including the release of the MLA decoding core FlashMLA specially built for NVIDIA Hopper GPUs, the EP communication library DeepEP, and the FP8 GEMM (Universal Matrix Multiplication) calculation library DeepGEMM. DeepSeek is expected to release new open source technology on the X platform around 9 a.m. today.
According to public information, for the whole year of 2024, OpenAI sales revenue will be approximately US$3.7 billion, a year-on-year increase of more than 1700%. It is expected that by 2025, OpenAI’s annualized revenue will increase to US$11.6 billion, of which 75% of the revenue comes from users ‘ChatGPT Plus service subscriptions. It is internally estimated that OpenAI revenue for the full year of 2029 will reach US$100 billion, equivalent to Nestlé’s current annual sales.
