OpenAI is reportedly working on two closely linked AI projects: "Strawberry" and "Orion." One is aimed at stronger reasoning, especially in math and programming. The other is described as a future flagship language model meant to outperform GPT-4.
The connection between them matters. Strawberry is not only being discussed as a possible chatbot feature. It may also help produce training data for Orion, giving OpenAI a way to improve the next generation of models with data generated by a more capable reasoning system.
What Strawberry is reportedly built to do
According to The Information, citing two people involved in the project, OpenAI might release a chatbot version of Strawberry as early as this fall, possibly as part of ChatGPT. The report also notes that it is uncertain whether Strawberry will launch this year.
Strawberry is described as a model designed for problems that current systems struggle with. Its focus includes previously unseen math problems and programming tasks that require optimization. The source also says the model's stronger logic could help it handle language-related challenges more effectively when it has enough time to "think."
That framing suggests Strawberry is less about instant response speed and more about deliberate problem-solving. In plain terms, the reported goal is to make an AI system that can work through harder tasks instead of only producing fluent answers quickly.
The source gives one example from internal demonstrations: Strawberry reportedly solved the New York Times word puzzle "Connections." That example is narrow, but it points to the kind of reasoning OpenAI may be targeting: pattern recognition, constraint handling and language understanding in the same task.
Why agent-based AI is part of the picture
Strawberry could also become a foundation for more advanced AI systems that do more than generate text or code. The source says the model may support systems capable of taking action.
Internal OpenAI documents reportedly describe plans to use Strawberry models for autonomous internet searches. In that version of the idea, the AI would be able to plan ahead and conduct in-depth research rather than simply respond to a single prompt.
That would make reasoning especially important. An AI system that searches, plans and acts needs to decide what to look for, how to sequence steps and when it has enough information. The source does not say how such systems would be released, but it does make clear that Strawberry is being discussed in connection with more autonomous workflows.
Reuters reported that OpenAI has already tested an AI internally that scored over 90 percent on the MATH benchmark, a collection of math mastery tasks. The source says this is likely Strawberry, and adds that Strawberry has also been presented to national security officials, according to The Information.
A smaller version may be the public product
If Strawberry is released, The Information says it would be a distilled version of the original model. Distillation is described in the source as a way to deliver similar performance with less computational power.
The source also notes that OpenAI has used this technique for GPT-4 variants since the original model was released in March 2023. That detail is important because it suggests a familiar pattern: a larger or more expensive model can inform a more practical version that users may actually access.
The reported approach resembles the "Self-Taught Reasoner" (STaR) method introduced by Stanford researchers. That method aims to improve the reasoning abilities of AI systems. Former OpenAI chief researcher Ilya Sutskever, who has since founded his own startup focused on secure super AI, is said to have provided the idea and basis for Strawberry.
How Strawberry may support Orion
Orion is described as OpenAI's GPT-4 successor. The source says OpenAI is working on the flagship language model with the goal of outperforming GPT-4.
Strawberry's role may be to generate data for Orion. That is a significant distinction: Strawberry may be useful not only as a model people interact with, but also as a tool for building a stronger model behind the scenes.
The source says the combination of Strawberry and high-quality synthetic data could potentially reduce errors in Orion compared to previous models. It also notes that research shows high-quality data is key for capable and efficient AI models.
OpenAI CEO Sam Altman mentioned at a May event that the company had gathered enough data to train its next model and was experimenting with synthetic data. In this context, Strawberry appears to fit into a broader training strategy: use a reasoning-focused system to help create better data, then use that data to improve a larger successor model.
OpenAI is not alone in the math race
The source also points to Google DeepMind's work on AI systems with advanced mathematical capabilities. Google DeepMind has developed AlphaProof and AlphaGeometry 2, which won silver at the International Mathematical Olympiad.
However, the source adds an important limitation: it remains unclear how well these models can be scaled and generalized. That is the broader question around systems like Strawberry as well. Strong performance on math, puzzles or programming tasks is useful, but the larger test is whether those abilities transfer across many kinds of real-world problems.
For now, Strawberry and Orion remain reported projects rather than fully public products. But the reported strategy is clear enough: OpenAI is exploring stronger reasoning, more autonomous research behavior and synthetic data as building blocks for a model beyond GPT-4.