Strawberry team

The Strawberry team is the OpenAI reasoning project team associated with o1. Its work brought longer inference-time reasoning into public use: the September 2024 release included o1-preview and o1-mini, while the main o1 model was still in training. The team joined OpenAI DevRel for a launch-era meetup.[1][4]

o1’s central technical distinction was spending more computation on reasoning before answering a question. In its September 2024 account, Latent.Space described additional reasoning tokens billed to users, with the raw reasoning chain hidden rather than necessarily appearing as a longer visible answer. In its September 12, 2024 research announcement, OpenAI reported that “o1 performance smoothly improves with both train-time and test-time compute.”[1][5]

The initial products represented different sizes and emphases: o1-preview was the larger released model, while o1-mini was described as a STEM-reasoning-focused distillation. These were distinct from the main o1 model still being trained. OpenAI’s September 12 product announcement said ChatGPT Enterprise and Edu users would receive access to both released models the following week.[1][6]

Before the public o1 release, Strawberry was an internal OpenAI project name. In July 2024, Reuters reporters Anna Tong and Katie Paul reported that the project had formerly been called Q* and was already regarded inside OpenAI as a breakthrough.[4]

In a November 25, 2024 discussion with Fireworks CEO Lin Qiao, swyx characterized the Strawberry team’s position as holding that its reinforcement-learning and chain-of-thought work could not be replicated by repeated open-source model calls. Qiao agreed that quality scaling was shifting from training to inference, but said Fireworks was taking a different approach built on existing models. Swyx disclosed that he advised Fireworks.[3]

The same discussion exposed a practical consequence of longer reasoning. Swyx said he had used Preview and Mini in production since launch and wanted control over their thinking budget. He described o1 as deciding that budget itself, and proposed letting users specify a duration or dollar amount when they knew how difficult a task would be.[3]