OpenAI’s Secret Project “Strawberry” is Out - Blog

OpenAI’s secret project “Strawberry” is out in the world and you can get a taste of it. OpenAI is resetting the counter on its models, starting afresh with o1-preview and o1-mini. These two new models are now live for paying users of ChatGPT with increased reasoning capabilities and much stricter limits.

What's going on here?

OpenAI's releasing a new series of AI models called OpenAI o1 that are designed to reason through complex tasks.

‍

What does this mean?

Right now, we have access to two models:

o1-preview: An early version of the upcoming full o1 model. It's designed to tackle complex reasoning tasks more effectively than previous models.
o1-mini: A smaller model with enhanced reasoning capabilities, suitable for tasks that require less computational power but still benefit from improved reasoning.

Paying users of ChatGPT can access o1-preview and o1-mini, but with stricter usage limits compared to GPT-4o. o1-mini gets 50 messages per day and o1-preview has a quota of 50 messages per week. For API users, tier 5 users can use these models (subject to rate limits). Common features like file and image uploads are not supported with o1 models yet.

The o1 models use a method called chain-of-thought (CoT), which means they internally work through intermediate steps before giving a final answer. This makes them slower (responses can take from a few seconds up to a couple of minutes) but significantly better at handling tough problems. ChatGPT summarizes these intermediate steps, but OpenAI keeps the raw reasoning hidden for safety and competitive reasons.

Why It’s Better:

Competitive programming: o1 scores in the 89th percentile, showcasing its strong problem-solving skills.
Academic expertise: o1 can surpass Ph.D.-level performance in difficult physics, biology, and chemistry problems.
Legal precision: In legal tasks, o1 doubles the accuracy of document revisions compared to GPT-4o.

Dan Shipper from Every highlights that o1 introduces "System 2 thinking," allowing AI to perform deep, deliberate reasoning similar to human thought processes. Professor Ethan Mollick notes that o1 can handle tasks that were previously impossible for AI, giving the impression of a new form of "agency" in its problem-solving approach.

Why should you care?

Taking time to think is crucial—not just for us, but for AI models too. The o1 models show that by allowing AI to spend more time on a problem, we can achieve better results on complex tasks.

While GPT-4o is still a fantastic tool for most everyday needs thanks to its speed and versatility, o1 really shines when it comes to tasks that require deep reasoning, planning, and multi-step solutions. This makes o1 perfect for things like detailed data validation, complex coding projects, or navigating tricky regulatory guidelines.

Here’s how to get started with o1:

Start with complex challenges: Use o1 for tasks that involve multiple steps or need detailed reasoning—areas where older models might struggle.
Provide clear instructions: Even though o1 is better at reasoning, giving detailed prompts can help the model understand the task more effectively.
Review outputs carefully: With o1’s deep reasoning, it’s important to check its outputs to ensure they’re accurate and relevant.

Our detailed guide to using o1 models at work is coming soon, stay tuned 👀.

‍