OpenAI announces GPT-5, claiming the model demonstrates human-level reasoning capabilities on complex tasks and introduces native agentic workflows.
OpenAI has officially released GPT-5, claiming the model achieves human-level reasoning on complex cognitive tasks. The release introduces native agentic capabilities, allowing the model to plan multi-step tasks, use tools, and learn from feedback without human intervention.
The model builds on the GPT architecture while integrating advancements from OpenAI's reasoning-focused o1 and o3 series. GPT-5 can now maintain context across significantly longer conversations and handle complex multi-step problems with remarkable accuracy.
"GPT-5 represents a fundamental shift in what AI can accomplish," said Sam Altman, OpenAI's CEO. "It's not just better at answering questions—it can actually reason through complex problems and execute multi-step plans."
On benchmark tests, GPT-5 scored 94% on graduate-level reasoning problems, matching or exceeding human experts in many domains. Perhaps more importantly, the model shows improved calibration, acknowledging uncertainty rather than hallucinating answers.
The release includes GPT-5 API access for enterprises, with pricing starting at $15/million input tokens for the base model. A specialized coding variant, GPT-5-Code, has also been released for developer applications.
OpenAI has implemented enhanced safety measures, including improved refusal capabilities and better detection of potentially harmful requests.
