NeuralStackOpenAI released the o3 reasoning model on July 22 2026. The model supports a 4 million token context window and achieves 92 percent on GPQA and 87 percent on MATH-500. It also includes native tool use for code execution and web search.
Training completed on July 10 using a cluster of 120000 H100 GPUs. OpenAI stated inference costs start at 8 dollars per million tokens for input. The release includes an API tier for developers with rate limits of 500 requests per minute.
Background context shows OpenAI has iterated on reasoning models since o1 in late 2024. o3 incorporates synthetic data generated from internal verification systems. The company trained on 18 trillion tokens total.
Competitors including Google and Anthropic have similar projects in testing. OpenAI positioned o3 as the first model to exceed human expert performance on several graduate-level benchmarks.
Why this matters
The launch accelerates the shift toward agentic AI systems that can plan multi-step tasks. Enterprises can now integrate longer context for legal and scientific workflows without chunking documents.
Forward-looking analysis indicates o3 will influence pricing across the industry. OpenAI plans a smaller o3-mini variant for mobile deployment by September 2026.