OpenAI has introduced two new models, o3 and o4-mini, designed to enhance the reasoning abilities of ChatGPT across various user tiers.

The o3 model handles advanced coding, math, and visual analysis. It handles multifaceted queries effectively and makes 20% fewer major errors compared to its predecessor.

The o4-mini model is optimized for cost-effective and rapid processing, excelling in math and coding tasks.

Both models can process images and access ChatGPT tools like web search and Python for richer, more accurate responses.

🔗 Source: OpenAI

🧠 Food for thought

1️⃣ Evolution of AI reasoning shows diminishing returns may not apply to “thinking time”

OpenAI’s development path with the o-series challenges the conventional wisdom about diminishing returns in AI improvement.

While most AI advancements face efficiency plateaus after initial gains, the new models demonstrate continuous performance improvements when given more computational resources for reasoning 1.

The o3 model achieves an 83% score on the International Mathematics Olympiad qualifying exam compared to GPT-4o’s 13%, showing dramatic improvements through extended reasoning capacity 1.

The results suggest that allowing AI systems to “think longer” continues to yield benefits without hitting a ceiling, potentially representing a fundamentally different scaling dynamic than traditional model size increases.

2️⃣ AI reasoning models reveal dramatically different design philosophies

The landscape of reasoning-focused AI models shows companies prioritizing different tradeoffs based on distinct strategic visions.

OpenAI’s o-series models are significantly slower but more capable in complex reasoning tasks, operating at about 1/30th the speed of GPT-4o but with vastly improved performance on mathematical and coding problems 1.

DeepSeek’s R1 model, developed at a reported cost of only $6 million, demonstrates that competitive reasoning capabilities can be achieved without OpenAI’s level of resources, suggesting different approaches to efficiency 2.

These divergent approaches create a spectrum of options: OpenAI’s models excel in deep reasoning with high accuracy but at increased cost and latency, while others prioritize speed and efficiency at the expense of some reasoning depth 3.

Recent OpenAI developments