OpenAI just released o3-mini and o3-mini-high, two new models focused on STEM reasoning tasks. While their performance improvements are notable, what really catches my attention is the inevitable flood of memes playing on that ‘high’ designation.
Let’s look at what these models actually do. o3-mini costs 93% less than o1 and beats it at PhD-level science questions and competitive coding. The cost per million tokens is $1.10, which still makes it about 7x more expensive than GPT-4o mini, but the performance boost might justify that for many users.
The ‘high’ version takes this even further. You can select different reasoning effort levels – low, medium, and high – to balance between speed and accuracy. This flexibility is genuinely useful, even if the name is destined to spawn endless drug-related jokes.
These models support all the usual developer features: function calling, structured outputs, streaming, and dev messages. They work across the Chat Completions API, Assistants API, and Batch API. Currently, they’re rolling out to select developers and paid users, with free users getting access through the “Reason” option.
The performance numbers are solid. o3-mini matches o1 on the toughest reasoning challenges and scores 99.2% on Frontier Mass. That’s better than both o1-mini and o1.
One interesting point is the safety angle. o3-mini apparently has some jailbreaking capabilities that challenge safety evaluations. OpenAI claims they’ve addressed this with “deliberative alignment,” but we’ll see how that holds up.
Compared to competitors like DeepSeek, which charges $0.55 per million input tokens, o3-mini is more expensive.
While everyone else will be busy making memes about o3-mini getting ‘high,’ I’m more interested in its practical applications. The combination of improved performance and lower costs makes it a compelling option for STEM-focused applications, even if we have to endure an endless stream of 420 jokes along the way.
