OpenAI o3 and o3-mini—12 Days of OpenAI: Day 12

OpenAI is announcing two new reasoning models—o3 and o3-mini—positioned as a step-change in performance on coding, math, and general reasoning benchmarks, while also expanding public safety testing before any broad release. The company is not launching them for general users yet, but it is opening applications for researchers in public safety testing starting immediately, with additional external access planned for o3-mini and eventually o3.

On the capabilities front, o3 posts strong results across technical benchmarks. In software-style coding tasks (SweetBench Verified), it reaches 71.7% accuracy, reported as more than 20% better than o1. On Codeforces competition coding, o3 is described as reaching an ELO near 2727 under aggressive test-time compute. For math, o3 scores 96.7% on the AIME benchmark (vs. 83.3% for o1) and 87.7% on GPQ Diamond, a PhD-level science/math benchmark (vs. 78% for o1). OpenAI also highlights a “hard mode” math test—Epic AI’s Frontier Math—where most systems score under 2% accuracy; o3 is reported to exceed 25% under aggressive settings, a notable jump on what’s framed as the toughest available math benchmark.

A centerpiece of the announcement is o3’s performance on ARC Prize’s Arc AGI benchmark, a long-running test designed to measure learning of new skills from examples rather than memorization. ARC Prize says its Arc AGI benchmark has been unbeaten for five years, and it reports o3 achieving a new state-of-the-art score of 75.7 on a semi-private holdout set under low compute, with 87.5 under high compute. The company ties this to a milestone: the high-compute score is described as surpassing a human-performance threshold (around 85%), and OpenAI says it has not previously seen a model clear that bar on this benchmark.

Alongside o3, OpenAI is rolling out o3-mini as a cost-efficient reasoning option. With “adaptive thinking time” exposed in the API (low, median, and high reasoning effort), o3-mini is presented as delivering coding performance that scales with more compute while remaining far cheaper than top-tier reasoning models. In live demos, o3-mini can generate and execute Python code via a local code-generation-and-execution workflow, and it can run a self-evaluation script against a hard GPQ dataset quickly using low reasoning effort. OpenAI also reports that o3-mini supports key API features such as function calling and structured outputs, aiming to make the reasoning models practical for developers.

Finally, OpenAI pairs the capability push with safety work. External safety testing access is opened for o3-mini starting today, with o3-mini planned for broader availability around the end of January and full o3 shortly after. The company also introduces “deliberative alignment,” a safety technique that uses model reasoning to better determine when to reject or review prompts, improving performance on rejection and refusal-related benchmarks. The overall message is clear: stronger reasoning models are arriving, but the rollout is gated by expanded safety evaluation and tighter alignment methods.