OpenAI uses secret Strawberry model to brew its next flagship Orion

Current AI models are great at writing emails and summarizing meetings, but give them a novel math problem or a convoluted programming task and they fall apart.

OpenAI is trying to fix that with two new models behind the scenes. According to The Information, the company is developing a reasoning model codenamed Strawberry and a next-generation flagship named Orion. OpenAI has reportedly even demoed Strawberry for U.S. national security officials.

Buying time to think

Strawberry is built around enhanced logical reasoning. Instead of spitting out the first plausible token, it takes extra time to think through problems it has never encountered before.

In internal demonstrations, Strawberry solved the New York Times word puzzle Connections. Reuters previously reported that an internal OpenAI system scored over 90 percent on the rigorous MATH benchmark, which is widely believed to be Strawberry.

Internal OpenAI documents also outline plans to use Strawberry for autonomous web research, allowing the AI to plan multi-step actions and conduct deep dives independently.

The underlying approach reportedly mirrors Stanford's Self-Taught Reasoner (STaR) method, focusing on letting models bootstrap their own reasoning loops. Former chief scientist Ilya Sutskever reportedly provided the initial ideas and foundation for the project before leaving.

Distillation for users, synthetic fuel for Orion

According to The Information, OpenAI might launch a chatbot version of Strawberry as early as this fall, possibly bundled inside ChatGPT, though a 2024 launch remains uncertain.

If it arrives, users won't get the raw, compute-heavy monster. OpenAI plans to release a distilled version that delivers comparable reasoning performance while burning significantly less compute, a playbook the team has honed since GPT-4 arrived.

Meanwhile, Orion is designed as the direct successor to GPT-4. With the open web running out of fresh, human-generated training text, Strawberry is being used to generate high-quality synthetic training data to teach Orion, specifically to cut down on hallucinations and reasoning errors. Sam Altman hinted at this shift back in May, confirming OpenAI had enough raw data and was aggressively experimenting with synthetic data.

Why it matters

We think this signals a definitive end to the brute-force era of language models. Simply throwing more web scrapes into pre-training is no longer delivering the leaps it used to; frontier labs are now shifting to test-time compute and self-improving reasoning loops.

For everyday users and developers, Strawberry is far more interesting than another slightly faster chatbot. If the distilled version makes it into ChatGPT, programmers tackling brittle architectural problems or researchers doing heavy synthesis will finally get an assistant that can plan before it speaks, making it worth testing the minute it drops.

Across the industry, this puts heavy pressure on Google DeepMind. DeepMind proved AI could score silver at the International Mathematical Olympiad with AlphaProof and AlphaGeometry 2, but scaling those specialized systems into general products remains an open question. OpenAI is betting that a general reasoning engine can both power user-facing agents and solve its internal training bottleneck.

However, training Orion on Strawberry-generated synthetic data looks like a high-wire act. It is a clever way around data exhaustion, but if Strawberry harbors subtle biases or logical blind spots, those flaws risk being baked into Orion at a fundamental level. Furthermore, pitching Strawberry directly to national security officials shows OpenAI knows autonomous, planning-capable agents are crossing straight into geopolitical scrutiny.