OpenAI wraps 12-day sprint by unveiling o3 reasoning models

OpenAI saved its biggest punch for the grand finale of its 12-day launch marathon.

The company officially unveiled o3, its next-generation reasoning model family, which includes both the flagship o3 and a smaller o3 mini. If you are wondering what happened to o2, OpenAI skipped the name entirely to sidestep trademark headaches with the European telecom giant O2.

The catch? Neither model is open to the public just yet.

Smashing benchmarks with adjustable brainpower

According to OpenAI, o3 comfortably beats o1 across math and science benchmarks. It set higher marks on AIME 2024, a competition-level math test, and GPQA, an evaluation packed with PhD-level questions in biology, physics, and chemistry.

More notably, o3 achieved a new state-of-the-art score on ARC-AGI. While OpenAI explicitly noted that the model has not reached AGI, the jump in abstract reasoning indicates it is creeping closer.

Alongside the heavyweight, OpenAI showed off o3 mini, which introduces three distinct reasoning effort modes: low, medium, and high. Basically, you decide how long the model gets to think.

At the low setting, o3 mini matches the performance of o1-mini on the Codeforces coding benchmark. Crank it up to medium or high, however, and it rivals or even beats the full-sized o1, a trend OpenAI says holds across other evaluations.

Rollout timeline and safety guardrails

You will have to wait a bit to actually use them. Sam Altman wrapped up the livestream by noting that o3 mini is slated to ship at the end of January, with the full o3 arriving sometime after.

In the meantime, OpenAI is opening external safety testing for the first time. Security researchers have until January 10 to apply for early access through an online form.

To keep reasoning models under control, the team also introduced deliberative alignment. It is a training framework that teaches reasoning models explicit, human-written safety guidelines, forcing them to reason through those rules before outputting an answer.

Why it matters

We think o3 signals that test-time compute is now the main battleground in frontier AI. Pumping compute into inference rather than just training runs continues to yield aggressive benchmark gains, putting immense pressure on rivals like Google and Anthropic.

For developers and power users, the three-tiered reasoning effort in o3 mini is the most practical breakthrough. Instead of waiting an unpredictable amount of time for a response, teams can budget latency and cost depending on whether a task needs a snap decision or deep reflection.

OpenAI played this smart by hyping future capability while buying time for external vetting. Still, deliberative alignment relies on the model accurately evaluating human safety rules during its hidden reasoning steps. As these systems get significantly smarter, verifying that reasoning actually stays aligned will be OpenAI's hardest test yet.