OpenAI Unveils o1: The Hyped 'Strawberry' Model Is Finally Here

OpenAI has announced the release of its new AI model, o1, also referred to by insiders as “Strawberry.” The new model is designed to handle more complex tasks and goes beyond previous capabilities by offering enhanced reasoning skills. While it promises significant advancements, it comes with a higher cost and slower performance compared to its predecessor, GPT‑4o.

Kylie Robison from The Verge reports that o1 is the first in a series of planned “reasoning” models aimed at answering more complex questions faster than humans. Alongside o1, OpenAI is also launching a smaller, less expensive version called o1‑mini. Despite being in its early stages, marked as a “preview,” o1 represents a shift towards human‑like artificial intelligence, especially excelling in writing code and solving multi‑step problems.

Access to o1‑preview and o1‑mini starts today for ChatGPT Plus and Team users, with Enterprise and Edu users getting access next week. Eventually, all free users of ChatGPT will have access to o1‑mini, although no definite release date has been provided. Developer access to o1 is quite expensive, costing $15 per 1 million input tokens and $60 per 1 million output tokens, which is significantly higher compared to GPT‑4o.

The new model's training process differs from its predecessors. According to Jerry Tworek, OpenAI’s research lead, o1 utilizes a new optimization algorithm and a specific dataset tailored for it. While previous GPT models were taught to mimic patterns, o1 utilizes reinforcement learning, a method involving rewards and penalties to teach the model to solve problems independently. This approach includes a “chain of thought” process to more accurately address complex queries step‑by‑step.

The accuracy of o1 is notably improved, with fewer instances of hallucination—where the model generates incorrect or nonsensical information. However, this problem isn't entirely resolved, as admitted by OpenAI. A key feature of o1 is its advanced problem‑solving ability in specialized areas like coding and math. The model has demonstrated remarkable performance, such as scoring 83% on a qualifying exam for the International Mathematics Olympiad compared to GPT‑4o's 13%.

OpenAI claims that o1 could potentially perform on par with PhD students in benchmark tasks across various scientific disciplines in future updates. Despite its impressive capabilities, o1 has limitations. It doesn't perform as well as GPT‑4o on factual knowledge, and it lacks the ability to browse the web or handle files and images. Nonetheless, OpenAI views o1 as introducing a new class of AI capabilities, indicating a shift in their approach to naming models, starting from “1.”

An interesting aspect of o1 is its user interface, designed to show the model's thought process as it solves problems. During a demonstration, the model was seen tackling a complex age‑related puzzle by breaking it down into steps and presenting its reasoning process. This step‑by‑step problem‑solving ability provides an illusion of human‑like thought, making the AI's operations appear more transparent and comprehensible.

Despite these advancements, Tworek highlights that OpenAI does not equate the model’s thought process with human thinking. The design choice aims to demonstrate the model’s deeper engagement with problems rather than superficial pattern recognition. This feature reflects OpenAI’s ambition to push the boundaries of AI towards more autonomous, decision‑making systems.

OpenAI's progress with the o1 model aligns with its larger goals of achieving more advanced AI capabilities that can someday lead to breakthroughs in fields like medicine and engineering. McGrew, OpenAI’s chief research officer, believes that the focus on reasoning is crucial for moving towards human‑like intelligence. Although the o1 model is still in its infancy, it paves the way for future developments in AI, emphasizing the critical role of reasoning in solving complex problems.

As OpenAI seeks to raise more funding at a staggering $150 billion valuation, continued research breakthroughs are vital for maintaining its momentum. The reasoning capabilities introduced with o1 could eventually lead to the development of autonomous agents capable of making decisions and performing tasks on behalf of users. However, the current model remains expensive and slower for developers, underscoring the challenges ahead.