“The word "strawberry" contains 3 occurrences of the letter ‘r’,” replied o1, OpenAI’s latest AI offering which has been embedded in ChatGPT on a preview basis. o1’s response is significant because the earlier iterations of ChatGPT—like GPT-4 and GPT-4o—struggled to count the number of ‘r’s in the word "strawberry."

Incidentally, o1 has been earlier addressed in the media by either Q-Star or Strawberry, as reported by The Secretariat.And after months of speculation, OpenAI released two iterations of the model—o1 and its cost-efficient counterpart o1-mini—on September 12.

But o1’s capabilities aren’t limited to just counting Rs or Ss in a word. It can apparently ‘think.’

Developed For More Complicated Tasks

What we’ve always expected and wanted from ChatGPT—quick answers—o1 does the exact opposite, mostly. It takes its own sweet time, processing everything slowly and methodically.

OpenAI said in a blog post that what sets o1 apart from the rest of its models—like GPT-4, Meta’s LlaMa and Google Gemini—is its ability “to spend more time thinking” before it responds. 

For example, to answer a text-based prompt, ChatGPT (running on GPT-4) would take perhaps 3-4 seconds to generate an answer. The o1 reportedly takes up to a minute to generate answers. OpenAI has likened this to how humans think before they respond.

You can observe the o1 model’s ‘thinking’ process by clicking the dropdown button next to it while it generates an answer and it also tells you how much time it spent ‘thinking’ about the answer. However, despite this visibility, does o1 really ‘think’ like a human? It appears not.

Human thinking involves consciousness, emotions, intuition, and experience, while AI models like o1 process data, recognize patterns, and predict outcomes based on the vast datasets they’re trained on.

While AI models can ‘mimic’ certain aspects of human reasoning—like making decisions—they lack self-awareness, and the ability to understand context.

Some might argue that AI models can ‘think’ since their structure is loosely inspired by the human brain. Neural networks in a human brain are like a series of algorithms designed to recognize patterns, interpret data, and make decisions or predictions, similar to how humans make decisions. 

Machine learning models are built on neural networks, which mimic the way neurons in the brain communicate.

o1’s Chain-Of-Thought

OpenAI has claimed that the o1 can “reason through complex tasks and solve harder problems than previous models in science, coding, and math.”

What o1 does is follow "chain-of-thought" programming, meaning it breaks down tasks step by step, whether solving a math problem or helping you create a Sudoku puzzle. Along the way, it also double-checks its answers, significantly reducing the chances of producing incorrect or hallucinated responses.

“We have found that the performance of o1 consistently improves with more reinforcement learning (train-time compute) and with more time spent thinking (test-time compute),” said the company.

According to the stats published by OpenAI, o1 ranks in the 89th percentile on Codeforces, and is among the top 500 in the USA Math Olympiad qualifiers. It’s also reportedly surpassing PhD-level accuracy in science benchmarks.

In comparison to other large language models, the big change with o1 is that instead of focusing most of the compute power on training the model beforehand or fine-tuning it after, more power is used during inference—when the model is actively solving a problem. 

o1 doesn’t need to memorize everything like other models. Instead, it uses tools in real time to figure out answers. This shift means it’s more efficient at reasoning when it’s actually being used.

Safe to say, o1 should be used to answer the big questions and not simpler ones that GPT-4 can handle with ease. The o1 model is thorough but overdoes it at times. When a 50-word paragraph would have sufficed for a question like—’how many states are there in the United States?,’ it can give an 800-word detailed explanation.

“o1 is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it,” said OpenAI CEO Sam Altman in an X post.

There has been some speculation if o1 is the successor to the GPT-4 model, but that seems unlikely. In an interview with the Wired, chief technology officer Mira Murati clarified that o1 is merely complimentary to the current models OpenAI has. She also said that GPT-5 is likely to also include the reasoning technology in o1.

The o1-preview and o1-mini models are available to ChatGPT Plus and Team users, with plans to extend access to Enterprise and Edu users in the future. OpenAI says it will add new features like browsing, file and image uploading.