Experience OpenAI's o1-preview Model: A Step-by-Step Guide to Getting Started

Despite months of speculation surrounding its development, OpenAI's launch of Project Strawberry last week came as a surprise to many analysts, who anticipated that the model would not be ready for weeks, if not later this fall.

The new o1-preview model, alongside its o1-mini variant, is now accessible for use and evaluation. Here’s how you can gain access.

OpenAI has introduced a preview of o1, a groundbreaking series of AI models designed to enhance their reasoning capabilities before generating responses. These models excel in tackling complex tasks and solving advanced challenges in domains such as science, coding, and mathematics.

What is o1?

OpenAI’s ambitions for artificial general intelligence (AGI) are well known, and Project Strawberry (now rebranded as “o1”) represents a significant step towards that vision. As the inaugural model in a new line prioritizing reasoning, it is crafted to “spend more time thinking before responding,” according to an official announcement. This methodology allows the model to effectively reason through intricate tasks and tackle more difficult problems than previous iterations in fields like science, coding, and mathematics. The models are designed to emulate human-like reasoning, enabling them to refine their thought processes, experiment with different strategies, and learn from their mistakes throughout training. OpenAI claims that o1-preview can perform comparably to Ph.D. students in subjects like physics, chemistry, and biology, achieving favorable outcomes in benchmark assessments in these disciplines. In addition to its prowess in scientific fields, o1 excels in coding and mathematical challenges, scoring 83% on an International Mathematics Olympiad (IMO) qualifying exam—where GPT-4o managed just 13%—and landing in the 89th percentile in a Codeforces competition against human participants.

What about o1-mini?

o1-mini is a streamlined version of the standard o1-preview model, reportedly operating at 80% lower costs than its larger counterpart. This makes it particularly effective in tasks relating to coding analysis and generation.

Is o1-preview available for testing?

Yes, the o1-preview models were launched on September 12, exclusively for ChatGPT Plus and Teams subscribers. Enterprise and Educational users will gain access starting the following week.

How secure is o1 against misuse?

According to reports, o1 has been developed with enhanced safety measures. OpenAI has established a new safety training program that utilizes the model’s advanced reasoning skills to ensure tighter adherence to safety and alignment protocols. Notably, in testing, while GPT-4o scored a mere 22 (out of 100) in resisting jailbreak attempts, the new o1 model achieved an impressive score of 84.

How can I access o1-preview?

Currently, the newly released o1-preview is available exclusively to paying subscribers. To try it out, you will need a $20/month Plus subscription. Click on the Upgrade Plan radio button in the left-hand navigation pane and follow the on-screen instructions to enter your payment details. Once your subscription is active, you can select either o1-preview or o1-mini from the model picker toggle on the left side of the ChatGPT homepage. Please note that access is limited even for subscribers, with a weekly cap of 30 messages for o1-preview and 50 messages for o1-mini. OpenAI has indicated that o1-mini will eventually be available to free-tier users, but a specific date for this rollout has not yet been established.