• OpenAI’s o3 achieves extraordinary breakthroughs in reasoning, coding and mathematics, far surpassing both human and AI benchmarks, including a 1200% improvement in math performance and an ARC-AGI score signaling AGI progress.

• With rapid progression from o1 to o3 in just three months, o3 exemplifies a shift in AI scaling through inference compute, leveraging chain-of-thought reinforcement learning (RL) to drastically shorten innovation cycles.

• By extending beyond single-task mastery to versatile problem-solving across complex domains, o3 demonstrates significant strides toward AGI, bridging the gap between specialized AI and generalized intelligence.

OpenAI Day 7-12 AI Releases Summary

Sources: OpenAI, Counterpoint Research

1. O3’s grand mastery in coding, math and science tops humans

To call o3 merely another state-of-the-art (SOTA) model from OpenAI would be an understatement; its reasoning capabilities represent a groundbreaking leap forward, far surpassing all existing models and going well beyond incremental improvements.

Coding

SWE-bench Verified is an agent-focused evaluation benchmark based on real-world software engineering challenges, such as GitHub issues. Below is a graph comparing the existing top-performing models.

Source: OpenAI

The achievement of a 71.7% score by o3 is extraordinary, and a direct 20% jump like this is unprecedented and game-changing. This leap mirrors the progress from GPT-4o, released as recently as May 2024, to Gemini 2.0 Flash and Claude Sonnet 3.5, models not even built specifically for SWE-tailored tasks.

Beyond model-to-model comparisons, o3 also outperforms 99.7% of human coders worldwide. As shown below, o3 has achieved an Elo score of 2727 on Codeforces, placing it among the top 200 competitive programmers globally — ahead of even OpenAI’s chief scientist, who holds an Elo of 2665. In contrast, DeepMind’s AlphaCode 2, released in December 2023 and specifically trained on Codeforces, reached the 87th percentile — a commendable expert-level milestone at the time. With o3, however, OpenAI has elevated the game, producing a true Codeforces Grandmaster.

Sources: OpenAI, Codeforces

Math and science

Starting with GPQA Diamond, the toughest reasoning benchmark on PhD-level science, o3 achieves an impressive 87.7%, far surpassing the average 70% scored by human PhDs in their fields of expertise.

On the most challenging math benchmark, Frontier Math, which can have problems so complex they can take professional mathematicians hours or days to solve, o3 delivers not a mere 20% improvement but an astonishing 1200% leap over the previous SOTA model.

Source: OpenAI

2. Inference-compute scaling is 4-8 times faster than pretraining paradigm

O3's true breakthrough, beyond its superior performance, lies in the rapid progression from o1 to o3 in just three months. This marks an unprecedented pace in scaling inference compute, driven by the new paradigm of chain-of-thought (CoT) RL, which enables o3 to acquire new skills on the fly — far outpacing the traditional pretraining cycle of introducing new models every 1-2 years.

This approach is emerging as one of the most highly anticipated trends in 2025, with forecasts pointing to significantly shorter development cycles for frontier models across the AI industry.

3. AGI is nearer

In the past, RL has produced super-intelligent AI systems for narrowly defined tasks. Typical examples include AlphaGo which defeated the world champion of Go, and AlphaStar which outperformed top pro players in StarCraft II.

These achievements, while groundbreaking, were confined to specific tasks with clearly defined reward functions. In contrast, o3 represents a significant advancement by extending RL capabilities to a broader spectrum of complex and nuanced tasks, such as advanced mathematics, software engineering and science, as we analyzed above. These fields lack simple reward structures, making the application of RL more challenging. Therefore, o3's innovation lies in its sophisticated reward engineering, enabling it to learn and adapt across multiple domains that OpenAI prioritizes.

AGI is considered to be the ability to efficiently acquire new skills and solve open-ended problems, beyond just performing specific tasks. François Chollet's 2019 ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence) benchmark stands as the leading measure of AGI progress, focusing on problems that are easy for humans but difficult for AI.

Source: ARC-AGI

The picture above showcases o3's clear lead in the ARC-AGI score, reflecting its exceptional ability to generalize and acquire new skills across diverse challenges. This progress elevates o3 from a task-specific model to a versatile problem solver, bringing us closer to the dawn of the AGI era.