Strawberry: OpenAI’s New Reasoning Model

Strawberry is the code name for OpenAI’s latest and most advanced reasoning model, officially known as OpenAI o1. The o1 model is distinguished by its ability to engage in deep thought, requiring more time to ponder problems before providing an answer, similar to a human. This enables it to tackle complex issues in fields like science, mathematics, and programming.

In a demonstration video, OpenAI showcased a stark contrast between o1 and its predecessors, GPT-4 and GPT-4O. When prompted in ChatGPT to “write a grammatically correct sentence without using any letter more than once,” o1 paused for 39 seconds before responding with “go fix my bed.” This is a simple task that previous models struggled with due to their reliance on predicting the next word (token) based on the preceding ones. o1, on the other hand, employs a chain-of-thought approach, breaking down complex tasks into smaller steps and solving them sequentially before generating a final response.

The model is also designed to be safe, with safeguards in place to ensure that its responses are helpful and avoid generating harmful or illegal content.

OpenAI offers two versions of o1: o1-preview, a preview version for testing and improvement, and o1-mini, a smaller, faster, and more cost-effective version designed for code generation. The company plans to make o1-mini freely available.

ChatGPT Plus subscribers can access both models, with current usage limits of 30 requests per week for o1-preview and 50 for o1-mini.

OpenAI asserts that o1 is comparable to a doctoral student, capable of providing in-depth and logical responses that result from careful thought and analysis. It has the potential to be valuable in strategic planning and other fields requiring deep, multi-step reasoning.