With the launch of o3-pro, let’s speak about what AI “reasoning” really does

On Tuesday, OpenAI introduced that o3-pro, a brand new model of its most succesful simulated reasoning mannequin, is now obtainable to ChatGPT Professional and Crew customers, changing o1-pro within the mannequin picker. The corporate additionally lowered API pricing for o3-pro by 87 p.c in comparison with o1-pro whereas reducing o3 costs by 80 p.c. Whereas “reasoning” is beneficial for some analytical duties, new research have posed elementary questions on what the phrase really means when utilized to those AI methods.

We’ll take a deeper have a look at “reasoning” in a minute, however first, let’s look at what’s new. Whereas OpenAI initially launched o3 (non-pro) in April, the o3-pro mannequin focuses on arithmetic, science, and coding whereas including new capabilities like internet search, file evaluation, picture evaluation, and Python execution. Since these instrument integrations sluggish response instances (longer than the already sluggish o1-pro), OpenAI recommends utilizing the mannequin for advanced issues the place accuracy issues greater than velocity. Nonetheless, they don’t essentially confabulate much less than “non-reasoning” AI fashions (they nonetheless introduce factual errors), which is a major caveat when searching for correct outcomes.

Past the reported efficiency enhancements, OpenAI introduced a considerable value discount for builders. O3-pro prices $20 per million enter tokens and $80 per million output tokens within the API, making it 87 p.c cheaper than o1-pro. The corporate additionally lowered the worth of the usual o3 mannequin by 80 p.c.

These reductions handle one of many principal issues with reasoning fashions—their excessive value in comparison with normal fashions. The unique o1 value $15 per million enter tokens and $60 per million output tokens, whereas o3-mini value $1.10 per million enter tokens and $4.40 per million output tokens.

Why use o3-pro?

In contrast to general-purpose fashions like GPT-4o that prioritize velocity, broad information, and making customers be ok with themselves, o3-pro makes use of a chain-of-thought simulated reasoning course of to dedicate extra output tokens towards working by advanced issues, making it typically higher for technical challenges that require deeper evaluation. Nevertheless it’s nonetheless not excellent.