OpenAI, the company behind ChatGPT, has unveiled its latest innovation: a new family of AI models under the name OpenAI o1, code-named “Strawberry.” Available in two versions, o1-preview and o1-mini, these models are designed to take a giant leap forward in AI’s ability to reason and even fact-check itself.
Unlike its predecessor, GPT-4o, the o1 model can spend more time thinking through its responses, making it more thoughtful in how it tackles complex tasks. But with this improvement comes a hefty price tag, and some limitations, especially for those hoping for more immediate functionality.
o1: What’s New?
The OpenAI o1 family of models is built to handle more complex reasoning tasks than previous models. While o1 can’t yet browse the web or analyze files like GPT-4o, it’s still impressive in its ability to break down multi-step problems and generate more accurate responses. Plus, it has image analysis features, although these are currently disabled for further testing.
For now, users will need to be either ChatGPT Plus or Team subscribers to access the model. Educational and enterprise clients will get access next week.
While o1-preview and o1-mini are now available, these models have some restrictions. There’s a weekly message limit (30 messages for o1-preview and 50 for o1-mini), and the models come with a premium price: $15 per 1 million input tokens and $60 per 1 million output tokens, which makes it three to four times more expensive than GPT-4o.
OpenAI has announced that o1-mini will eventually be available for free users, but there’s no confirmed release date yet.
A Model That Can Fact-Check Itself
What truly sets o1 apart is its unique approach to reasoning. The model can now “think” more deeply before responding, which helps avoid some of the mistakes generative AI models tend to make. This more deliberate thought process is perfect for tasks that require careful analysis and multiple steps — think analyzing a lawyer’s inbox for privileged emails or brainstorming a marketing strategy.
According to OpenAI research scientist Noam Brown, o1 is trained using reinforcement learning, which encourages the model to spend more time working through questions. Through a system of rewards and penalties, the model learns to improve its reasoning over time.
OpenAI also used a specialized dataset for training o1, including scientific literature and “reasoning data,” making it especially well-suited for coding, data analysis, and science-related queries. As Brown explains, “The longer it thinks, the better it gets.”
Impressive Early Performance
Even before its public release, o1 has demonstrated remarkable improvements in problem-solving. In a qualifying exam for the International Mathematical Olympiad (IMO), a global high school math competition, o1 solved 83% of the problems, a vast improvement over GPT-4o’s 13% success rate. For coding challenges, the model ranked in the 89th percentile of participants in programming competitions, outperforming systems like Google DeepMind’s AlphaCode 2.
Ethan Mollick, a professor at the Wharton School of Business, has been testing o1 over the past month. He found it to be particularly strong when faced with multi-layered tasks. In one test, the model accurately solved a crossword puzzle, even inventing a new clue.
But no AI is perfect, and o1 is no exception. Mollick noted that while the model can sometimes be slow, it can also get tripped up by simpler tasks — like playing tic-tac-toe. And despite its enhanced reasoning, o1 still suffers from hallucinations, or confidently making up false information.
A Competitive Race
OpenAI’s release of o1 is part of a larger trend in AI research. Other companies, including Google DeepMind, are also exploring ways to improve model accuracy and reasoning by allowing AI more time and resources to process information. This kind of innovation might seem abstract, but it has real-world applications, from legal analysis to programming help.
While OpenAI has a head start, its competition is moving quickly. One interesting note: OpenAI opted not to display o1’s “chains of thought” in real-time in ChatGPT, partly to keep a competitive edge. Instead, users will see a model-generated summary of the reasoning process.
What’s next? OpenAI says it’s planning future iterations of o1, with the goal of creating models that can spend hours, or even days, solving particularly challenging problems.
For now, we’ll have to wait and see how o1 performs in the hands of more users and whether the high cost is worth the added depth of reasoning it offers. One thing is clear, though — the race to create smarter, more thoughtful AI is just getting started.
