OpenAI o3

On ToTheMoon, OpenAI o3 is the model the hosts used through 2025 to work out what a reasoning model means in daily work. The questions are practical: is the reasoning real or a chain of substituted concepts (episode 038), why does bundling tools into one window fail to fix hallucinations (episode 055), and how to choose between o3, GPT-4o and Deep Research when no strongest model exists (episode 061).

The first conversation, at the end of 2024, was wary: o3 was available almost only to researchers, and nobody was highlighting safety benchmarks for reasoning models — if there were something to boast about, they would be boasting about it (episode 038). Once the model shipped, the assessment became concrete: o3 is interesting for integrating a large context, search, memory and images in one mode, yet there is no simple winner — Grok may be stronger in code, and the interface promises more certainty than the model gives (episode 055). By May one host admitted it was hard to move off o3 (episode 059), and episode 061 laid out a split: GPT-4o for everyday questions, o3 for what matters in business.

By summer 2025 the tone sobers. An expensive subscription does not guarantee a better answer, and o3 stumbles on the two-sisters riddle from SimpleBench (episode 062). o3 Pro answered very slowly at first, and then users noticed o3 itself getting worse — the familiar cycle of a model launching strong and slipping (episode 063). Episode 067 states that “o3 Pro is no longer the best model on the market,” and in the ChatGPT agent test the first run beat o3, o3 Pro and Deep Research (episode 068).