OpenAI GPT-4.5: Boring and vicious

OpenAI has finally released the legendary GPT-4.5 (internal codename Orion), which may be the most bizarre major update in the history of AI - it is both boring and extremely vicious.

poverty

OpenAI’s pricing this time is simply insane!

The fee is $75 for every 1 million input tokens, and as high as $150 for output tokens.

That is, GPT-4.5 costs 30 times more than GPT-4o and 25 times more than Claude 3.7 Sonnet! In fact, it costs 10-100 times more than other models on the market!

This is not a pricing, this is simply a price that discourages customers from buying !

OpenAI itself admits that the high operating costs of GPT-4.5 may make them reconsider whether to open the API in the long term.

That’s right , this is a “ go online and lie flat ” model.

boring

Logically speaking, a model with such a sky-high price should have breathtaking performance, right?

The result of GPT-4.5 is just to tell everyone: the traditional pre-training model has come to an end!

Its pre-training computation is 10 times that of GPT-4! The result? The benchmark is only 5% better than 4o. In terms of code testing, it is worse than Claude 3.5, but not as good as Claude 3.7.

To sum it up in three words: extremely embarrassing! Expensive, slow, and not good enough.

Sam Altman also confidently said that GPT-4.5 brought "subtle improvements". Although it was not outstanding in hard-core reasoning areas such as mathematics and code, it reached new heights in understanding human intentions and emotional responses.

Translation: We spent a huge amount of computing power, but only created an AI that is more "empathetic".

Sam himself admitted that "it's not good at scoring, but it has ideas" - emmmm, does this sound like an excuse for hallucinations?

Internal testers reported that GPT-4.5 is "warmer, intuitive, and natural" and has "stronger aesthetic intuition and creativity."

In human terms:

Our model has changed from T-type personality to F-type personality . We understand your feelings but cannot solve your problems? ‍♂️

Extremely evil

The real "extreme evil" thing is not the model itself, but OpenAI's future development path.

Altman claims that GPT-4.5 will be the "last non-thinking chain model", which sounds harmless, but its actual impact may be extremely bad.

As the current de facto industry leader, OpenAI's strategic choices will affect the entire AI ecosystem.

The thought chain model is indeed a good thing, but not all scenarios are suitable for the thought chain.

Thinking chain has disadvantages: slow speed, high cost, and low efficiency. In the real world, enterprises and ordinary users may need efficient and practical APIs, not every question requires the model to ponder for several minutes before answering.

The chain of thought should be an option, not the only path.

Not all tasks require a chain of reasoning - that's as ridiculous as asking you for directions and the other person insisting on telling you the entire process from the formation of the earth's plates to urban planning.

This "full inference" trend will lead to other vendors following suit: API costs soar, small and medium-sized enterprises are squeezed out of the market, response time is increased, user experience is reduced, simple tasks are complicated, and resources are wasted.

What’s more frightening is that this approach may further concentrate AI technology in super-large technology companies because only they can afford such computing costs.

the truth

Ilya Sutskever once said bluntly: "We have reached the peak of data, and traditional pre-training methods are about to end." The performance of GPT-4.5 seems to confirm this.

My personal interpretation is that OpenAI is trying to pave the way for its high-cost new route by artificially creating technical bottlenecks. The market needs a hundred flowers to bloom - some scenarios require careful consideration, and some scenarios require rapid response. Forcibly directing all demands to a single technical route is not what an industry leader should do. How the development direction it represents will reshape the entire AI industry.

Efficiency and intelligence should be two parallel development paths, not an either-or situation.

Finally, I would like to say something to Altman: Be a good person!