Late-night big news! OpenAI releases its big move: GPT-4.5 is officially launched! But the price is too high...

This morning, OpenAI dropped a "technical bomb" without warning - the GPT-4.5 research preview was officially launched. What kind of black technology is hidden in this model, which is officially called "the closest to human conversation experience"? Is it really worthy of the "sky-high price" label? Let's find out.

As the latest iteration of the GPT series, GPT-4.5 continues OpenAI's "violent aesthetics" - using more data and computing power to "feed" a stronger model. However, there are signs that the benefits gained by expanding the scale of data and computing are flattening.

In multiple AI benchmarks, GPT-4.5 underperformed Chinese AI companies DeepSeek, Anthropic, and OpenAI’s own new “reasoning” model. OpenAI  ’s focus this time was on human interaction and reliability:

1. Emotional empathy

According to official tests, when a user enters "I feel miserable after failing an exam", GPT-4.5 will respond to emotional needs first ("This sounds really hard") and then provide practical suggestions. In contrast, the responses of GPT-4o and o3-mini are more "tool-like".

2. The rate of hallucinations has been greatly reduced

When generating factual content, the probability of GPT-4.5 "making up nonsense" dropped from 61.8% of GPT-4o to 37.1%. OpenAI particularly emphasized that this improvement is of great significance to high-risk scenarios such as legal consultation and medical assistance.

3. Hidden creative talent

In the SVG code drawing test, only GPT-4.5 successfully generated a recognizable unicorn image (see figure). This ability to convert abstract instructions into concrete output is eye-opening for designers and copywriters.

It’s worth noting that GPT 4.5 is not a “reasoning model” — a new class of models from companies like OpenAI, DeepSeek, Anthropic, and others that generate “thought chains” or stream-of-consciousness-like chunks of text that reflect on their own assumptions and conclusions in an attempt to catch errors before providing a response to the user. GPT-4.5 is still more like a traditional LLM.

Still, according to OpenAI co-founder and CEO Sam Altman in a post on social platform X, GPT-4.5 is: "The first model that made me feel like I was talking to a thoughtful person. There were several times I leaned back in my chair and was amazed at the truly useful advice I got from the AI."

Despite its impressive technology, GPT-4.5's pricing strategy has caused an uproar:

-Subscription fee : Currently only available to ChatGPT Pro users at $200/month, Plus users will have to wait until next week to get it

- API prices are sky-high : processing 1 million words of text costs $75/$180 for input/output (compared to $2.5/$10 for GPT-4o)

OpenAI CEO Sam Altman admitted on social media: "We are stuck in a GPU shortage." In order to alleviate the computing power crisis, the company is urgently deploying tens of thousands of graphics cards, but the high-price strategy is difficult to change in the short term.

Impact on enterprises:

- ✅ High-frequency scenarios such as customer service and content generation may reduce costs due to improved accuracy

- ❌ Small and medium developers say they can't afford it, and some teams have turned to the more cost-effective Claude 3.7 or Gemini 2 Pro

In each generation of GPT before GPT-4.5, scale expansion has brought significant leaps in performance in a number of areas, including mathematics, writing, and programming. OpenAI said that the scale growth of GPT-4.5 has indeed given it "deeper world knowledge" and "higher emotional intelligence."

Third-party testing revealed the "biased" phenomenon of GPT-4.5:

Interestingly, OpenAI’s own employees seemed to have made a mistake - the technical documents leaked before the press conference deleted the key description of "computing efficiency increased 10 times", and the official explanation was "inaccurate description". This operation made netizens joke: "It turns out that even big companies cannot escape PPT failures."

Although the price is discouraging, three types of users may consider it:

1. Content creators

- Automatically generate copywriting consistent with the brand tone

- Quickly generate marketing plans based on image materials (supports multiple file uploads)

2. Corporate Services Department

- 7x24 hours emotional customer service to reduce customer complaints

- Automatically organize meeting minutes + extract key points

3. Technology Early Adopters

- Develop customized AI tools through API

- Experience the interactive design that is currently closest to "human dialogue"

-Supporters : "GPT-4.5 proves that there is still potential for large models. Its 'empathy' may open up new tracks such as psychological counseling and educational companionship." (AI analyst Andrew Curran)

- Skeptics : "The investment increased 1,000 times, but the MMLU score only increased by 5 points? This cost-effectiveness makes people doubt that the technology has reached its ceiling." (Anonymous AI company founder)

- OpenAI's ambition : The official defines this release as a "research preview", suggesting that the model may be optimized through user feedback in the future, and even multimodal functions such as voice and video may be integrated.

GPT-4.5 is like a "liberal arts student with a rapidly improving emotional intelligence" - good at chat creation, but a little struggling in the hardcore field. OpenAI plans to eventually combine its GPT series models with the "o" reasoning series, starting with GPT-5, which will be launched later this year.

GPT-4.5, which was reportedly extremely expensive to train, had multiple delays, and failed to meet internal expectations, may not win the AI ​​benchmark crown on its own, but OpenAI likely sees it as a stepping stone to more powerful technology.