- OpenAI o1 is a new model series (starting with o1-preview and o1-mini) focused on real-time reasoning, not just pre-training.
- o1 outperforms GPT-4o in STEM tasks, scoring 83% on IMO qualifiers vs. GPT-4o's 13%, and ranks in the 89th percentile on Codeforces.
- o1 uses chain-of-thought reasoning natively, making it slower but significantly more accurate on complex tasks.
- o1 is much harder to jailbreak, scoring 84/100 on adversarial safety tests vs. GPT-4o's 22.
After months of speculation from ChatGPT fans, OpenAI’s latest Strawberry LLM release has dropped – and it's not called GPT-5.
Previously referred to with cryptic, intriguing codenames like Q* and Strawberry, the latest model series has finally settled on the moniker OpenAI o1.
The biggest update with OpenAI's new model is its enhanced reasoning skills. OpenAI explained that o1 has been trained to spend more time thinking than previous models, bringing it closer to human intelligence.
What is OpenAI o1?
OpenAI o1 is the latest series of large language models released by OpenAI on September 12, 2024, currently comprising two models: o1-preview and the o1-mini.
The biggest difference between o1 and the company's previous models is its advanced reasoning. While it’s not yet released in full, the preview and mini models already blow GPT-4o out of the water on tests of math, science, and coding.
OpenAI o1 models
The September release included two models, the o1-preview and the o1-mini. They're the first of a series of models that will continue to be released as OpenAI refines their new LLM.
The difference? The o1-mini model is smaller and 80% cheaper than the preview model. It’s built for tasks that require advanced reasoning, but not broader knowledge. It’s perfectly suited for tasks that involve coding or mathematics.
How smart is OpenAI o1?
OpenAI has touted a list of STEM benchmarks that show off o1’s reasoning abilities, including:
- A similar performance to PhD students in benchmark tests on physics, chemistry, and biology.
- Placing in the top 500 students in the US qualifier for the USA Math Olympiad.
- Ranking in the 89th percentile in Codeforces, a competitive coding test.
You can read more about o1's reasoning abilities in OpenAI's research release.
What is chain of thought reasoning?
The o1 models use chain of thought reasoning, a longer and more thorough way of breaking down requests.
If the o1 model is given a prompt, it won’t answer immediately - hence the long wait time. Instead, it will reason through each of the steps, carefully considering each piece of information and its implications before deciding on the next course of action. It won't provide an answer until it has thought through the entire series of steps required in the ask.
How is o1 different from GPT-4o?
1) Reasoning abilities
At the center of its general intelligence is o1’s new ability to reason. “Maybe the most important areas of progress will be around reasoning ability,” Altman shared with Gates. “Right now, GPT-4 can reason in only extremely limited ways.”
Reasoning is notoriously difficult. Even for humans. And OpenAI o1 is the first model to claim it.
The o1 models are able to reason in real time, rather than rely on pre-training data. This is why the new model is better at science, math, and coding tasks than previous OpenAI models.
2) Harder to jailbreak
With safety concerns on the rise as LLMs grow in popularity and power, security has been a key focus of OpenAI’s latest development. The company partnered with the U.S. and U.K. AI Safety Institutes while developing the o1 series, as well as worked with the American government to establish their due diligence.
As a major step forward, the o1 series is far harder to jailbreak – bypass safety measures – than previous models.
On one of their hardest jailbreaking tests, o1-preview model scored 84 out of 100, compared to a dismal 22 score from GPT-4o.
3) New naming convention
While its name isn't the most exciting thing about the new OpenAI LLM, it is an intentionally meaningful change.
OpenAI o1 is the first model to cast off the 'GPT' moniker, and that's because the company claims it's the first phase of a brand new 'reasoning paradigm', whereas the older models were part of a 'pre-training paradigm'.
The new model spends time reasoning in real time, rather than relying on its pre-training data.
4) Better at STEM problem-solving tasks
With better reasoning comes better math skills.
Both o1 and GPT-4o were asked to complete a qualifying exam for the International Mathematics Olympiad. GPT-4o solved 13% of problems, while o1 solved 83%.
5) Longer wait time
Reasoning in real time takes longer than referencing training data and generating a response. If you ask a question to OpenAI o1-preview compared to other models, you'll be waiting significantly longer.
However, with the ability to outsource reasoning, it's a small price to pay. The speed of the o1 models will likely improve as the next models in the series are released.
Who can use o1?
As of September 12, ChatGPT Plus and Team users are able to access o1 models in ChatGPT.
OpenAI announced that they’ll make o1-mini available to free users, although no date has been pinpointed.
The current weekly rate limits are 30 messages for 01-preview and 50 for o1-mini, though they will soon increase.
What should I use o1 for?
The enhanced reasoning capabilities of o1 are especially useful for solving complex problems in math, science, and coding. As OpenAI explains:
Limitations of OpenAI o1
As a preview, this model doesn’t yet have all the capabilities of GPT-4o. If you’re looking to use an LLM to browse the web for information, or you want to upload files or images, you’ll need to stick to GPT-4o until later models of o1 are released.
How to prompt OpenAI o1
OpenAI’s prompting suggestions have changed for o1 compared to their previous models, due to its enhanced reasoning.
Keep your prompts simple. It’s a smart model, and doesn’t need as much guidance as the GPT-4 series. That means avoid any chain of thought input – the model is already reasoning internally.
Build GPT-powered AI agents
What if your AI agent automatically synchronized with every OpenAI update?
Botpress is a completely open and extendable AI agent platform. Our stack allows developers to build chatbots and AI agents with any capabilities, across any workflow.
The only platform that ranges from low code set-up to endless customizability and extendability, Botpress allows you to automatically get the power of the latest GPT version on your chatbot – no effort required.
Start building today. It’s free.
FAQs
What does “o1” stand for? Is there a naming logic for future models (e.g., o2, o3)?
"o1" likely stands for "OpenAI 1" and marks the beginning of a new model series so yes, you can probably expect o2, o3, and beyond as future versions.
Why did OpenAI move away from the “GPT” naming convention?
OpenAI says o1 represents a shift from the old "pre-training paradigm" to a new "reasoning paradigm," so the new name reflects that change in direction.
Is OpenAI o1 the successor to GPT-4 or an entirely new product line?
While it builds on GPT-4’s progress, o1 kicks off a new generation of models focused on real-time reasoning.
Can o1 handle multi-language inputs better than GPT-4?
OpenAI hasn’t said much about multilingual support yet, but since o1 is designed to reason better, it may eventually outperform GPT-4 in that area too.
When will o1 be generally available for free-tier users?
o1-mini is supposed to roll out to free-tier users soon, but OpenAI hasn’t shared an exact date yet.
.webp)
