Forget About GPT-5, Let's Meet OpenAI's Latest "Strawberry" and "Orion" Models First

GPT-5 has been long-awaited, but instead, the mysterious Strawberry and Orion models emerged, placing OpenAI again at the center of attention.

According to the latest media report, OpenAI's enigmatic Strawberry (formerly known as Q*) project will be launched as early as this fall. Concurrently, the new large-scale model Orion, developed using synthetic data from Strawberry, is expected to be introduced at the beginning of next year.

Previously, OpenAI has hinted several times at the launch of the Strawberry model. ChatGPT posted a strawberry emoji on X in early August, seemingly implying the Strawberry large-scale model. Sam Altman also recently posted a picture of a strawberry on social media.

What are the Mysterious "Strawberry" and "Orion"?

Strawberry is rumored to have a powerful reasoning ability, capable of handling complex problems. It is not limited to answering technical questions but can handle complex problems, such as solving programming and mathematical problems, which has always been a challenge in traditional conversational AI.

In addition, the Strawberry model will not only provide training data for Orion but may also be integrated into ChatGPT in a simplified version to enhance the reasoning ability of GPT-4.

Orion, on the other hand, is the next-generation LLM that surpasses GPT-4, significantly reducing hallucinations or errors during the process. The importance of Strawberry technology to Orion is self-evident. Researchers have revealed that Strawberry can generate high-quality training data for Orion. The more correct complex reasoning examples Orion is exposed to through training data, the better the learning effect.

OpenAI's founder and CEO, Sam Altman, has emphasized that the key to future AI development will revolve around reasoning ability.

Due to the limitations of reasoning ability, current AI technology has not been widely applied in mathematically intensive industries such as aerospace and structural engineering. Therefore, optimizing the mathematical reasoning ability of large models has become a necessary condition for the popularization of AI applications and a necessary path for AI companies to further enhance their profitability.

The predecessor of the Strawberry model is the Q* algorithm model, which can solve tough scientific and mathematical problems. If an AI model masters mathematical abilities, it will have stronger reasoning abilities, and may even be able to solve mathematical problems it has never seen before or spontaneously solve transformation problems, rivaling human intelligence.

It is worth mentioning that the Strawberry model has scored over 90% in the MATH benchmark test, with an impressive overall performance.

Some netizens believe the Strawberry model will become a game-changer in the financial and other technology industries, completely solving complex calculations and reasoning methods.

From a product roadmap perspective, OpenAI may generate a simplified version of the original Strawberry model through distillation, integrating the improved reasoning ability into GPT. The simplified version aims to maintain the same performance level as the original model while being easier to operate and less costly.

Although this may make the answers more accurate, the time spent may be longer, which means that Strawberry may not be suitable for applications requiring instant responses. It is expected that ChatGPT users will be able to choose to enable or disable the Strawberry feature according to their needs.