An internal document and source detail OpenAI's Strawberry, formerly known as Q*, which aims to improve its AI models' reasoning capabilities dramatically
ChatGPT maker OpenAI is working on a novel approach to its artificial intelligence models in a project code-named “Strawberry,” …
Context & Ripple Effects
This report establishes Strawberry as an internal OpenAI effort centered on improving model reasoning, rather than a conventional product update. It provides the technical premise for later coverage of a selectable ChatGPT release and the eventual o1 preview.
The project also sits at the intersection of model development and strategic positioning: later reporting said OpenAI briefed US national-security officials on Strawberry and could use it to generate training data for Orion.
First-order effects
- OpenAI’s near-term model work becomes more explicitly organized around reasoning performance, creating a distinct capability track that later surfaced as o1 alongside a smaller o1-mini.
- Strawberry’s proposed role in generating training data ties its value not only to end-user answers, but also to improving a subsequent flagship model.
Second-order effects
- A reasoning-focused product track gives OpenAI a basis for segmenting access and model choice, as reflected in the later ChatGPT Plus and Team preview of o1.
- If reasoning systems can help produce training data for later models, progress in one model generation can compound into the development pipeline for the next, increasing the strategic importance of evaluation and deployment controls.
Third-order effects
- The episode points toward AI competition shifting from general-purpose chat models to differentiated reasoning systems that serve both users and internal model-building workflows.
- As advanced capabilities receive national-security attention, leading AI labs may face a more state-mediated path to deployment; the extent of that shift depends on whether such engagement becomes routine rather than exceptional.
The trend: Frontier AI labs are turning reasoning into a distinct product and infrastructure layer, with outputs increasingly feeding both customer-facing tools and future-model development.