OpenAI releases GPT-4.5: a "subtle upgrade" or "dissuasive" pricing?

High-profile AI giant OpenAI recently announced the release of a research preview of its latest flagship language model, GPT-4.5, once again generating a lot of attention from the tech community. The highly anticipated new model will initially be available to software developers and ChatGPT The opening of Pro subscribers heralds the start of a new round of AI technology competition.

GPT-4.5: Stronger Conversation Skills and Higher Resource Requirements

According to OpenAI in its official blog post, GPT-4.5 is the largest and most costly language model introduced by the company so far. It achieves significant improvements in dialog capabilities and aims to provide a more fluid and natural interaction experience. However, it's worth noting that more performance also means higher resource consumption; GPT-4.5 requires more GPU resources to run, which undoubtedly puts a higher demand on the computing infrastructure.

OpenAI claims that early tests show that the experience of interacting with GPT-4.5 more closely resembles natural human conversation. This is largely due to its broader knowledge base, more accurate understanding of user intent, and so-called "higher emotional intelligence". OpenAI believes that these improvements will make GPT-4.5 better at tasks such as writing, code editing, and solving real-world problems. In addition, OpenAI emphasizes that GPT-4.5 outperforms previous GPT-4o and o1 models in an illusion rate test, showing progress in generating accurate information.

Sam Altman, co-founder and CEO of OpenAI, also posted a positive review of GPT-4.5 on social media platform X. He said that his experience with GPT-4.5 made him feel like he was talking to a "thoughtful person" for the first time, providing valuable advice and even giving "excellent to surprising" advice. He said that his experience with GPT-4.5 made him feel for the first time like he was talking to a "thoughtful person" who was able to provide valuable advice and even give "amazingly good" answers. However, Altman also made a point of saying that GPT-4.5 is not known for its reasoning ability, and that its goal is not to outperform other models in benchmarking.

It is understood that the intelligence improvement of GPT-4.5 is mainly due to "unsupervised learning" (i.e., autonomous learning from massive data), which is different from OpenAI GPT-4o or GPT-4o, which focuses on inference ability. DeepSeek The R1 model is different. In terms of functionality, GPT-4.5 supports web search, can handle document and image uploads, and can use Canvas for writing and programming. However, GPT-4.5 does not yet support multimodal features such as voice mode, audio/video and screen sharing.

In terms of the user opening program, GPT-4.5 will initially be open to software developers and ChatGPT Pro subscribers. Then, it will be gradually opened to Plus and Team users in the following week, and then to Enterprise and Edu users in the following week.

GPT-4.5 Pricing Strategy Causes Controversy

However, while OpenAI officials are actively publicizing the performance improvements of GPT-4.5, there is another voice in the industry. One commentator suggested that the release of GPT-4.5 was perhaps the "strangest major update" in the history of AI development - "both boring and evil". This comment centers on GPT-4.5's pricing strategy.

Reviews have said that OpenAI's pricing is "insane": Input tokens cost $75 per million, and output tokens can cost up to $150 per million. This means that GPT-4.5 costs 30 times as much as GPT-4o, is Claude 3.7 Sonnet is 25 times more expensive! It has even been suggested that the GPT-4.5 is 10-100 times more expensive than other similar models on the market!

With such a high price tag, it doesn't seem too much of a stretch to use the term "exit price" to describe it. Even OpenAI itself has admitted that the running cost of GPT-4.5 is so high that it may need to reconsider whether to open the API in the long run. This slightly exaggerated statement reflects the industry's misgivings about the pricing strategy of GPT-4.5.

Performance Improvement "Subtle"? Bottlenecks in traditional pre-training models

Faced with such an expensive model, people naturally have high expectations for its performance. However, the actual performance of GPT-4.5 does not seem to bring "breathtaking" surprises. Comments indicate that the launch of GPT-4.5 may be more like a hint that the traditional pre-training model may have reached a bottleneck.

GPT-4.5 is said to have 10 times more pre-training computations than GPT-4, but benchmarks show a performance gain of only a "pitiful 5%" over GPT-4o. In code testing, GPT-4.5 outperforms Claude 3.5 but falls short of Claude 3.7.

"Embarrassing in the extreme! It's expensive, it's slow, and it's not good enough." One review used these three words to summarize the overall performance of GPT-4.5. Sam Altman tries to explain that GPT-4.5 brings "subtle enhancements," and that while mediocre in areas of hardcore reasoning such as math and code, it reaches new heights in understanding human intent and emotional response.

"We spent sky-high arithmetic power, but only made an AI that is more 'empathetic'." This interpretation may be slightly tongue-in-cheek, but it also points out the performance characteristics of GPT-4.5. According to feedback from internal testers, GPT-4.5 is "warmer, more intuitive, and more natural" and has "stronger aesthetic intuition and creativity." In other words, GPT-4.5 may be better at understanding and responding to human emotions, but the improvement in solving practical problems may be relatively limited.

"The last non-thinking chain model"? OpenAI's Strategic Turn and Potential Risks

The real "evil" of GPT-4.5 is not so much the model itself, but the future of OpenAI that it foreshadows. Altman declared that GPT-4.5 will be "the last non-chain-of-mind model". While this statement may seem innocuous, it could have a profound impact on the entire AI ecosystem.

As the current leader in the AI field, OpenAI's strategic choice will have a demonstration effect on the entire industry. The chain of thought model certainly has its advantages, but not all application scenarios fit with it. The chain of thought model has shortcomings such as slower speed, higher cost, and relatively low efficiency. In the real world, enterprises and ordinary users may need more efficient and practical API services, rather than all the questions need to be modeled "thinking" a few minutes to give the answer. Chain of Thought should be an option, not the only path.

"Not all tasks require a round-robin toss of reasoning chains- it's as absurd as asking for directions and having the person have to tell you the whole story from the formation of the earth's plates to the planning of cities." The review uses a vivid metaphor to point out the potential problems of "total reasoning." This trend, if followed blindly, could lead to a host of problems such as skyrocketing API costs, market exclusion of SMBs, longer response times, degraded user experience, complication of simple tasks, and wasted resources. More worryingly, this course of development could lead to a further concentration of AI technology in the mega-corporations, which are the only ones that can afford such huge computational costs.

All in all, OpenAI's release of GPT-4.5 may be a bit "subtle" in terms of performance improvement, but its high pricing strategy and declaration of "the last non-chain-of-thinking model" have undoubtedly triggered deep thoughts in the industry about the route of AI development. The first is to be a "good" one. The release of GPT-4.5 may be just the beginning of where AI technology will go in the future.