type
status
date
slug
summary
tags
category
icon
password
OpenAI has officially launched the research preview of GPT-4.5, currently the largest and most advanced GPT model. This model marks significant progress in unsupervised learning and reasoning capabilities, while also optimizing human-computer interaction and reducing hallucination.
GPT-4.5 is designed to be the apex of non-reasoning models, paving the way for future technological advancements, such as GPT-5. It excels in various domains, including writing, programming, and practical problem-solving, while significantly mitigating "hallucination," the generation of inaccurate or fabricated content.
GPT-4.5 is currently available to ChatGPT Pro users and developers, with a phased rollout planned for other user groups.
However, it is not a revolutionary upgrade, nor does it surpass reasoning models in specific areas. The release of GPT-4.5 has been met with some disappointment. For users, GPT-4.5 offers a smarter and more reliable chat experience, but its high cost and partial functionality gaps have sparked some discussion.
Furthermore, GPT-4.5 underperforms Claude 3.5 and even Deepseek V3 in certain benchmark tests.

Nevertheless, Sam Altman, CEO of OpenAI, holds a different perspective. He stated that GPT-4.5 is their "last generation of non-reasoning models," and future products will integrate more technologies such as reasoning capabilities.
.png?t=1a9b5dc8-d074-8044-87a9-c946b0804323)
“This is not a reasoning model, and it won’t break any benchmarks. This is a different kind of intelligence, and I have never experienced such magic. I’m really looking forward to everyone trying it!”
According to OpenAI's official introduction, GPT-4.5 is an exceptionally large and computationally intensive model. Its training utilized more computational resources and data than ever before, with a computational scale potentially 10 times that of GPT-4.
GPT-4.5 exemplifies the scaling of unsupervised learning through increased computational power and data volume, coupled with architectural and optimization innovations. Trained on Microsoft Azure AI supercomputers, GPT-4.5 has evolved into a model with broader knowledge and a deeper understanding of the world, resulting in reduced hallucination and enhanced reliability across various topics.
Key Features of GPT-4.5
Scale and Depth of Knowledge:
- GPT-4.5 is OpenAI's largest model to date, trained using more computational resources and data than previous models.
- Compared to GPT-4o, it possesses a deeper and broader "world knowledge," enabling it to comprehend more complex backgrounds and contexts.
- OpenAI states that the increased model scale reduces the need to "fabricate" answers, significantly diminishing hallucination.
Emotional Intelligence and Conversational Naturalness:
- GPT-4.5 demonstrates a breakthrough in "emotional intelligence," enabling it to better capture human intentions and emotions.
.png?t=1a9b5dc8-d074-805e-8c11-e153e9dae06b)
- User feedback indicates that GPT-4.5 exhibits more natural conversations compared to GPT-4o, particularly excelling in everyday queries, professional tasks, and creative writing (such as poetry).
General-Purpose Design:
- Unlike OpenAI's "o-series" reasoning models (e.g., o1, o3-mini), GPT-4.5 is not specifically focused on mathematical or scientific reasoning but is positioned as a general-purpose model.
- It is suitable for a wide range of applications, including writing, programming, and problem-solving.
Functionality Support:
- Supports real-time web browsing, file and image uploads, and ChatGPT's Canvas tool.
- Currently does not support Voice Mode, video, or screen sharing features.
Performance Metrics
OpenAI has provided several benchmark test results to showcase GPT-4.5's capabilities:
SimpleQA Test (a common-sense question answering benchmark developed by OpenAI):
- GPT-4.5 scored 62.5%, significantly higher than GPT-4o's 38.6% and o3-mini's 15%.
- Hallucination Rate: GPT-4.5 exhibited a hallucination rate of 37.1%, lower than GPT-4o's 59.8% and o3-mini's 80.3%.
.png?t=1a9b5dc8-d074-809c-94c3-c753d5e4c4f6)
MMLU Test (language understanding benchmark):
- Shows limited improvement compared to GPT-4o but exhibits a slight advantage in language tasks.
- Mathematics and Science Tests:
- GPT-4.5 underperforms reasoning models like o3-mini in these areas, indicating it is not focused on STEM (Science, Technology, Engineering, Mathematics) tasks.
Researchers point out that benchmark tests may not fully reflect the real-world experience of GPT-4.5. Users may perceive more significant improvements in non-quantifiable tasks such as writing and programming.
Comparison with Previous Models
Differences from GPT-4o:
- GPT-4.5 invested more resources in the pre-training phase, representing a scale jump comparable to the transition from GPT-3.5 to GPT-4o.
- It reduces hallucination, enhances writing ability and conversational fluency, but does not introduce groundbreaking new features.
Distinction from Reasoning Models (o1, o3):
- o-series models excel in chain-of-thought reasoning, suitable for mathematical and scientific problems.
.png?t=1a9b5dc8-d074-8012-b91f-c7b1697e5214)
- GPT-4.5 prioritizes instant responses and general-purpose utility, without focusing on reasoning as its core strength.
OpenAI CEO Sam Altman stated that GPT-4.5 is their "last generation of non-reasoning models," and future products will integrate more technologies such as reasoning capabilities.
Availability of GPT-4.5
GPT-4.5 is currently available in the ChatGPT Pro version and will be gradually rolled out in the future:
.png?t=1a9b5dc8-d074-80f6-9b37-ec147b0fc861)
- Next Week: ChatGPT Plus and Team plan users can access GPT-4.5.
- The Week After: Enterprise and Education plan users will gain access.
API Preview is now open:
- Supports file and image input
- Supports function calls, structured output, streaming responses
- Does not support Voice Mode, video, screen sharing (potential future additions)
Pricing Information
Input: $75.00 / 1M tokens
Cached input: $37.50 / 1M tokens
Output: $150.00 / 1M tokens
.png?t=1a9b5dc8-d074-807d-9d7b-d91764ba05db)
GPT-4.5 User Reviews: Mixed Reactions
User reviews of GPT-4.5 on platform X are polarized. Some users acknowledge its advancements in conversational naturalness, breadth of knowledge, and creative writing. Conversely, others express disappointment with its limited performance improvement, high pricing, and shortcomings in programming and reasoning tasks. A detailed analysis follows:
Positive Reviews
Conversational Naturalness and Creative Writing:
- Many users commend GPT-4.5's natural conversational flow, particularly highlighting its performance in creative tasks. For instance, one user noted its "strong intuition and interesting personality" when writing stories and poems.
- One user mentioned that while GPT-4.5's responses can sometimes be overly ornate and lengthy, adjusting prompts can tailor the style to individual needs.
Breadth of Knowledge and Image Understanding:
- User feedback suggests improvements in GPT-4.5's knowledge base and image processing capabilities. One user pointed out its ability to accurately process complex visual information and extract data, outperforming previous models.
- Furthermore, GPT-4.5's multilingual capabilities have been enhanced. Even users primarily focused on English content have noted stronger performance in non-English languages.
Humanized Improvements:
- Some tweets indicate that GPT-4.5 reduces instances of refusing to answer, provides more human-like replies, and adopts more standardized formatting, resulting in a more user-friendly overall experience.
Negative Reviews
Limited Performance Improvement:
- Multiple users believe that GPT-4.5's performance gains are not significant enough. One tweet stated that the "differences are subtle and not obvious" compared to GPT-4o, with progress being almost imperceptible in ordinary tasks.
- Other users complained about GPT-4.5's significantly slower response speed compared to GPT-4o, impacting daily usability.
Pricing Controversy:
- GPT-4.5's pricing has sparked widespread dissatisfaction. The API costs are $75 per million input tokens and $150 per million output tokens, significantly higher than GPT-4o's $2.5 and $10, respectively. Users question its cost-effectiveness, deeming it "too expensive."
- The ChatGPT Pro subscription fee of $200 per month has deterred many users, who express unwillingness to pay and prefer waiting for a free version or opting for more economical alternatives.
Insufficient Programming and Reasoning Capabilities:
- In programming and reasoning tasks, GPT-4.5's performance is underwhelming. Some users claim it underperforms Claude 3.7 Sonnet, particularly exhibiting a noticeable gap in coding capabilities.
- One user stated that after 30 minutes of testing, while GPT-4.5 could provide interesting suggestions, its assistance in solving practical technical problems was limited.
Questioning Scalability Strategy:
- Some users believe that OpenAI's approach of solely enhancing performance by scaling model size is nearing its limits. A tweet pointed out that GPT-4.5 validates the "diminishing returns of scaling," suggesting a need for more innovation in the future.
Neutral Reviews
Model Positioning:
- Some users argue that GPT-4.5 is positioned to provide natural conversations and broad knowledge rather than focusing on reasoning or programming, hence its mediocre performance in certain tasks is expected.
- Another tweet mentioned its improved linguistic nuance and emotional intelligence, but these advantages are difficult to quantify through tests and require practical usage to appreciate.
Wait-and-See Attitude:
- Some users advise against subscribing to ChatGPT Pro for now, recommending waiting for more comprehensive evaluations. They believe that the current performance improvements may not justify the high cost.
Official Introduction: https://openai.com/index/introducing-gpt-4-5/
- Author:KCGOD
- URL:https://kcgod.com/openai-releases-gpt-4.5
- Copyright:All articles in this blog, except for special statements, adopt BY-NC-SA agreement. Please indicate the source!