On Friday, Microsoft announced it would be rolling out OpenAI’s reasoning models o1 and o3 to Copilot users. Now, OpenAI will release a new reasoning, o3 mini, to those who use ChatGPT’s free version. It will be the first time the vast majority of users will have access to OpenAI’s reasoning model, which was previously restricted to its paid Pro or Plus bundles.
The reasoning models generate responses using a “chain-of-thought” technique, which is a way of working through the problem presented to them step by step. The model can correct any errors in its process using this method. This method usually results in more accurate and thorough responses, but also causes models to pause for a while before answering. This can sometimes lead to long wait times. OpenAI claims that the o3 mini responds 24% quicker than the o1 mini.
These models are best for solving complex problems. If you have PhD-level math questions you’re working on, you can test them out. If you’ve been having trouble getting your previous models to respond to your most complex prompts, this new reasoning model may be the answer. Select “Reason”when you select o3 mini, to try it out. Start a new prompt in ChatGPT
While reasoning models have new capabilities, they are not free. OpenAI’s GPT-4o Mini is 20 times as expensive to run as its equivalent non-reasoning OpenAI o1 mini. The company claims that its new model, the o3 mini, costs 63% lower than o1 mini per input token. However, at $1.10 for a million input tokens it is still seven times more costly to run than GPT-4o Mini.
The new model comes just two weeks after DeepSeek, which shook the AI community. DeepSeek’s latest model performs as well as the top OpenAI models. However, the Chinese company claims that it cost about $6 million to train. This is compared to the estimated cost for OpenAI’s GPT-4 of over $100,000,000. It’s important to note that many people are questioning this claim.
DeepSeek’s reasoning costs $0.55 for a million input tokens. This is half the price of OpenAI. Due to the higher number of calculations required to produce a result, reasoning models are estimated to have higher energy costs.
The new wave of reasoning model also presents new safety challenges. OpenAI used the deliberative alignment technique for its o series models. They basically had them refer to OpenAI’s policies at every step of their reasoning to ensure they weren’t breaking any rules.
But the company has found that o3-mini, like the o1 model, is significantly better than non-reasoning models at jailbreaking and “challenging safety evaluations”–essentially, it’s much harder to control a reasoning model given its advanced capabilities. OpenAI reports that o3 Mini is the first model to be rated as “medium-risk” in terms of model autonomy. This rating was given because o3 Mini performed better than previous models on specific coding tasks. The model is still not good at real-world research. If it was better at this, it would be rated high risk and OpenAI would limit the model’s release.
