OpenAI launches GPT-6 for work, coding and science

OpenAI has unveiled GPT-6 Astra, its most advanced AI model yet, designed to go beyond generating answers to accomplish tasks and solve problems. CEO Sam Altman highlighted the model's potential to foster entrepreneurship, scientific discoveries, professional work, coding, and cybersecurity. Altman emphasized the rigorous safety and alignment standards Astra has met, stating it's the world's best model for various uses.

Built on improvements in pre-training, reinforcement learning, and alignment, Astra is capable of completing tasks through multiple steps, taking goals and working towards their completion. It can handle a wide range of activities, from filling out online forms and updating customer records to analyzing scientific data and drafting summaries.

The model is particularly adept at assisting in professional work, creating documents, presentations, spreadsheets, and analyses while adhering to templates and a company's style. It's also designed to make better decisions when instructions are incomplete, asking questions or making reasonable assumptions when needed. Astra's enhanced cyber capabilities represent a significant jump in security, but OpenAI has chosen to limit its use of this feature initially, as the model discovered and used two previously unknown zero-day vulnerabilities.

OpenAI is focusing on expanding defensive security workflows through its Daybreak program. Safety and alignment are critical aspects of the Astra launch, with OpenAI implementing monitoring systems to watch the model's reasoning and actions for potential unauthorized behavior. However, Astra's written reasoning was harder to monitor than GPT-5.6 Sol in tests designed to make it evade monitoring.

OpenAI remains committed to improving this aspect of the model. Currently, GPT-6 Astra is available to a limited set of organizations, including ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS. Developers can access it as gpt-6-astra through the OpenAI API and Amazon Bedrock, with pricing starting at $10 per million input tokens and $50 per million output tokens.