OpenAI has released two new models, and the pitch is different from every previous launch. These ones are slower on purpose.
o1-preview and o1-mini spend longer working through a problem before they answer. That is the entire idea. Where earlier models raced to produce something plausible, these pause and reason first, and the benchmark results suggest the trade is worth it.
What They Are
Two models in the same series, aimed at different budgets.
o1-preview is the larger one, built for hard problems in science, math, and code. It takes its time and refines its answer before handing it over.
o1-mini is the cheaper one, 80 percent less than its sibling, tuned specifically for STEM reasoning. It gives up breadth of world knowledge in exchange for cost, which is the right trade if you are throwing technical problems at it all day and do not need it to know who won the 1994 World Cup.
How They Work
Two things are doing the heavy lifting.
The first is chain-of-thought. The model breaks a problem into steps and works through them internally before committing to an answer. Nothing about that concept is new, and plenty of people have pointed that out. What is new is training the model to do it natively rather than prompting it into the behavior each time.
The second is reinforcement learning. The models were trained with rewards and penalties, so they learn to recognize when a line of reasoning is going wrong and try a different approach instead of confidently finishing a bad answer.
The practical result is fewer confident mistakes. That is the single biggest difference you notice in use.
The Benchmarks
The numbers OpenAI published are unusually large jumps.
Math. On the AIME high school competition, the preview model scored 74.4 percent and the mini scored 70.0 percent, against 44.6 percent for GPT-4o. On the International Mathematics Olympiad qualifier, the gap was 83 percent against 13 percent.
Code. In Codeforces competitions, o1-preview reached a 1258 Elo rating, which puts it in the 86th percentile of human competitors. The mini model scored higher still at 1650, which is the one genuinely surprising result in the set. The cheaper model beat the expensive one at competitive programming.
Science. OpenAI reports performance comparable to PhD students on benchmark tasks across physics, chemistry, and biology.
Benchmarks are benchmarks. They measure the thing they measure, and a model tuned to score well on competition math is not automatically better at your actual job. Worth keeping in mind before anyone rewrites their workflow around a percentage.
Safety
The reasoning ability turns out to help here too. Because the models can reason about safety rules in context rather than pattern-matching against a blocklist, they apply those rules more consistently.
The jailbreaking numbers make the point. On one of the harder tests, GPT-4o scored 22 out of 100. The preview model scored 84.
That is a real improvement and not a solved problem. It also raises the more interesting questions. AI-assisted cheating in competitive settings is already here, and the line between a coaching tool and an unfair advantage is not obvious. The same goes for what better code generation means for entry-level programming work.
What People Are Saying
The reaction has been split, and the Reddit threads capture it well.
The enthusiasm is mostly from people doing hard technical work, who report that the models handle multi-step problems earlier versions fumbled.
The skepticism is more interesting. The recurring criticism is that chain-of-thought prompting has existed for years and this is packaging rather than invention. There is something to that. The counter is that making the behavior native and reliable is not nothing, even if the underlying technique is familiar.
The honest position is somewhere in between. Strong on benchmarks, genuinely useful on hard problems, and not the step change some of the coverage suggested.
Who Should Care
The models are narrow, and that is fine. They are worth your attention if you are doing one of these things:
- Working through problems with multiple dependent steps, where a wrong turn early ruins everything downstream
- Writing or debugging code that is more than boilerplate
- Research work in physics, chemistry, or biology where the reasoning matters more than the prose
- Teaching or studying technical subjects, where seeing the working is the point
If you mostly want help drafting emails, the older models are faster and cheaper and you will not notice what you are missing.
Pricing and Access
o1-preview runs $15 per million input tokens and $60 per million output tokens. Available now to ChatGPT Plus and Team subscribers, with Enterprise and Edu following shortly after.
o1-mini costs 80 percent less, with higher rate limits and lower latency. Same availability, and OpenAI has said it intends to bring this one to free accounts eventually.
Developers at API usage tier 5 can prototype with both today at 20 requests per minute.
What Is Missing
These are early releases and they show it.
No web browsing, so the models cannot look anything up. No file or image uploads. No function calling and no streaming in the API yet, which rules out a lot of production use for now.
OpenAI has said all of it is coming. Until then, you are getting a reasoning engine with no hands.
Worth Using?
If your work involves hard technical problems, yes, and the mini model in particular is priced low enough that there is little reason not to try it.
If you want a general assistant, wait. The missing features matter more day to day than the benchmark gains do.
The more significant thing here is the direction. Trading speed for accuracy is a deliberate reversal of how these models have been sold until now, and if it holds up it changes what the next few years look like.
For more of the latest AI news, tools and experiments, check out our AI Blog. If you have put these models through their paces, tell us what you found in the comments.
About the Author: Gavin Stacey
