These types of models are most effective in solving complex problems. Therefore, if you have a doctoral level math problem, you can try it if you are cracking down. Alternatively, if you have a problem with the appropriate prompt of the previous model to the most advanced prompts, try this new reasoning model. To try O3-mini, just select the reason when starting a new prompt in Chatgpt.
Progress models have new features, but cost. O1-mini of OPENAI is 20 times more expensive than the same non-rational model GPT-4O mini. The company, the new model, is 63 % cheaper than the O1-mini per input token, $ 1.10 per input token, and is 7 times more expensive than the GPT-4O mini.
This new model is right after the release of DeepSeek, which shakes the AI world two weeks ago. The new model of DeepSeek works like a top open model, but Chinese companies take about $ 6 million in training, in contrast to the estimated cost exceeding $ 100 million in Openai’s GPT-4 training. I claim. (It is noteworthy that many people are interrogating this claim.)
In addition, the price of DeepSeek’s inference model is $ 1 million and costs half the price of O3-mini, so Openai is still a way to reduce costs. The reasoning model is presumed to have much more energy cost than other types, given that there are many calculations required to generate answers.
The new waves of this reasoning model also bring a new safety issue. Openai uses a technique called Demiberative Alignment to train the O series model, and basically has referenced Openai’s internal policy at each stage of inference to confirm that it is not ignoring the rules. Ta.
However, the company has discovered that the O3-mini is much better than the O1 model, which is much better than the non-rational model in the Jail Break and the “challenging safety evaluation”. In essence, it is much more difficult to control the reasoning model considering advanced functions. O3-mini is the first model to score as a “medium risk” of the model’s autonomy. It states that it is better than the previous model of a specific coding task, and is likely to be self -improvement and acceleration of AI research. However, this model is still bad in real world research. If it gets better, it will be high risk, and Openai will limit the model release.
