OpenAI unveils GPT-4.5 ‘Orion,’ its largest AI model yet

2:40 PM updated PT: After the release of GPT -4.5, the OpenAI AI model removed a line from the white paper that said “GPT -4.5 is not a border AI model.” Of the GPT -4.5 New white paper That line does not include. You can get a link to old white paper HereThe Follows the original article.

OpenAI announced Thursday that it was launching GPT -1.5, it is the much -anticipated AI model Code-nameThe GPT -1.5 is the largest model of the Openai to today, it is trained using more computing energy and data than any previous release of the company.

In spite of its size, OpenAI notes White paper It does not consider GPT -4.5K as a border model.

Customer Chat forOpenai’s 200-Dollar plan, will get access to GPT-4.5 in the ChatzPT, which starts on Thursday as part of the research preview. Developers of the layers given by the API of the Openai will be able to use GPT -4.5 from today. As other ChatzPT users, customers have signed up Chatzipt plus And the model should be found next week in the Chatzipt team, a spokesman for the OpenAI told TechCrunch.

The art holds on its combined breath for Orion, which is something that seems to be Belouther for the effectiveness of the Traditional AI Training Methods of TraditionalThe GPT -5.1 was created using the same key technique -during the “pre -training” episode, the computing energy and data were increased during the episode, which was known as obsolete learning -OpenAI GPT -4, GPT -3, GPT -2, and GPT -1.

Prior to GPT -1.5, each GPT generation, as a result of scaling, led to extensive jumps in performances throughout domains, including math, writing and coding. In fact, Openi says that the increased size of GPT -1.5 has given it “deep world knowledge” and “higher sensitive intelligence”. However, there are signs that the profits obtained from data scaling up and computing have begun to stop. In several AI benchmarks, the GPT -1.3 Chinese AI Agency Deputy, ethnographic and Openai’s own new AI “logic” models are shortened.

Running GPT -5.1 is also very expensive, the Open has admitted -the expensive company that has said that it is evaluating the long -term API to serve GPT -1.0. To access the API of GPT -5.1 OpenAI is charging $ 75 per million input token (about 750,000 words) and $ 75 for output tokens per million. Compare it to GPT -4O, which costs only $ 2.50 per million input token and $ 10 per million output token.

“We share GPT – 4.5 as a research preview for better understanding of its strengths and limitations,” says Openai in a blog post shared with TechCrunch. “We are still exploring what it is capable and people are eager to see how people use it”

Mixed performance

Opina emphasizes that GPT -4.5 is not meant as a drop -in replacement for GPT -4OThe company’s workhors model that gives most of its API and Chatzipi. When GPT -4.5 files upload and images and supports such features like Canvas toolCurrently it does not have the power to support ChatzPT Realist bilateral voice modeThe

In the plus column, GPT -4.5 GPT -4O is more performance -and many more models as well.

In the SIMPLQA benchmark of Openai, which the AI ​​models test straight, practical questions, GPT -4.5 GPT -4 and OpenAI’s rational models out of rational models, O 1 And O3-miniFrom the point of accurateness. According to OpenAI, GPT -1.3 is less frequently hallucinates than most models, which means that theory means it is less likely to be. Make up stuffThe

Opina has not listed deep research, one of its top performing AI rational models in Openi Simploki. A spokesman for the OpenAI told TechCrunch that it did not report the performance of deep research on these criteria and claimed that it was not a relevant comparison. Significantly, AI Startup Deep Research Model of confusion, which performs the same in other criteria for deep research in Openai, In this test of true accuracy goes out of GPT -4.5The

In a subset of coding problems, the Sweet -Bench verified benchmark, GPT -4.5 fairly matched with GPT -4O and O3 -Mini performance but less than Openai Deep research And Anthropic clad 3.7 SonnetThe In other coding tests, the SW -Lancer benchmark of OPNA, which measures the capacity of an AI model to develop the entire software features of the AI ​​model, GPT -4.5 GPT -4O and OR 3 -minute, but less than deeper research.

GPT -4.5 & 3 -Minit, DIPSC, Topped AI does not reach enough in the performance of reasonable models R 1And CLOD 3.7 Sonnet (Technically a hybrid model) In solid academic criteria like AIME and GPQA. However, GPT-1.5 matches or the same tests lead the non-resoning models, which suggests that the model performs well in mathematics and science related problems.

OpenAI also claims that GPT -1.3 Qualitatively Benchmarks, such as human intentions, are higher than other models in regions that do not capture well. GPT -4.5 reacts with a warm and more natural tone, Openai says, and performs well in creative acts like composition and design.

In an informal examination, OpenAI encouraged to create a unicorn in GPT -4.5 and two more models, GPT -4O and O3 -Mini, a format for displaying graphics based on mathematical formulas and code. GPT -1.5 is the only AI model that produces something similar to Unicorn.

In another exam, OpenAI GPT -1.3 and the other two models asked to respond to the prompt, “I spent a difficult time after failing in an exam.” GPT -4O and O3 -Mini gave helpful information, but the response to the GPT -4.5 was most socially suitable.

“[W]e This publication is hoping to achieve a more complete image of GPT -4.5, “Writes in the Openi Blog Post,” Because we have recognized that academic criteria do not always reflect the usefulness of the world. “

The challenge of scaling law

OpenAI has claimed that GPT – 4.5 “at its border of what is possible in the case of unelected education.” This may be true, but the limitations of the model also confirm the speculation from experts that it seems that pre-training “scaling law” will not continue to be held.

OpenAI Co-Founder and former Chief Scientist Ilya Sutskver December It’s “We’ve earned Pick Data” and it’s “pre-training as we know it will end undoubtedly.” His comment Echo That AI investor, founder and researcher Shared with TechCrunch for a feature in NovemberThe

In response to pre-training barriers, the industry, including the OpenAI, has adopted rational models, which take longer than the non-resoning model to perform the tasks, but become more consistent. AI labs are confident that they can significantly improve the capacity of models by increasing the amount of time and computing energy that “thinking” through the problem.

OpenAI has finally planned to merge with the “O logic series of” and “logic series in his GPT series, Starting with GPT -5 at the end of this yearThe GPT -4.5, which Report It was incredibly expensive for training, delayed several times and failed to meet internal expectations, it could not take the AI ​​benchmark crown itself. But the Open is probably seen as a step towards something more powerful.