Reasoning Models Go Mainstream
OpenAI began the phased rollout of its o3 reasoning model to ChatGPT Plus ($20/month) subscribers this week, with access initially capped at 30 messages per day. The o3 model, which uses chain-of-thought reasoning before generating responses, shows particularly strong performance on mathematics, coding, and multi-step logical problems compared to GPT-4o.
What o3 Does Differently
Unlike GPT-4o which generates responses in a single forward pass, o3 allocates variable "thinking time" to reason through problems before answering. For simple requests this adds minimal latency; for complex problems, o3 may spend 30-90 seconds reasoning before responding. Users can observe the reasoning process in a collapsible "thinking" pane.
Performance Benchmarks
o3 scores 96.7% on the 2024 AIME (American Invitational Mathematics Examination), compared to GPT-4o's 73.2%. On SWE-bench (software engineering tasks), o3 resolves 71.7% of GitHub issues correctly versus 49% for GPT-4o. On GPQA Diamond (graduate-level science), o3 achieves 87.7%.
Rate Limiting and Pricing
The 30 message/day cap for Plus subscribers reflects genuine inference cost challenges — o3's extended thinking process is substantially more compute-intensive than standard generation. OpenAI has indicated it will raise the cap as it scales H100 and Blackwell GPU capacity. Pro subscribers ($200/month) receive 150 o3 messages per day at launch.
Competitive Response
Anthropic's Claude 3.7 Sonnet with extended thinking and Google's Gemini 2.0 Flash Thinking are the direct competitors. All three companies are converging on similar reasoning architectures, suggesting this approach has become the new baseline for frontier performance.