OpenAI launches GPT-6 Astra, called closest to AGI

OpenAI announced GPT-6 Astra, which the company described as its closest model yet to artificial general intelligence (AGI). The launch included a summary of technical and safety testing materials released by the firm.

OpenAI described the model as showing broader problem-solving across multiple domains and improved ability to apply reasoning and context over longer interactions compared with prior versions. The company positioned the model for use in complex research, advanced software development and other work that requires sustained cognitive effort.

The firm plans a phased rollout beginning with enterprise customers, selected partners and researchers while monitoring performance and safety. Access will be limited as OpenAI expands safety checks, red-team testing and usage controls to limit harmful or unintended behavior.

Training and alignment updates are intended to reduce hallucinations and improve reliability on ambiguous or unfamiliar queries. The release includes features to make outputs more interpretable for human reviewers and to provide clearer signals about confidence and uncertainty in responses.

Technical details highlight multimodal inputs and a longer effective context window, allowing the model to process text, images and structured data across prolonged sessions. OpenAI noted the model can refine its approach during extended interactions and transfer learning patterns from one domain to related tasks without retraining.

The company reported internal evaluations and external audits and will continue working with academic and industry groups to assess risks. OpenAI indicated new guardrails for content safety, data handling and user access, and that some high-risk use cases will remain restricted under its policy framework.

OpenAI plans to publish more technical materials and safety findings in the coming weeks to enable independent review. The company encouraged researchers and partners to take part in evaluation efforts and to report any unexpected behaviors during monitoring.

OpenAI acknowledged public debate about the definition and implications of AGI and clarified that GPT-6 Astra does not equate to human-level intelligence. The company described the release as reflecting progress in the model’s ability to generalize across tasks compared with earlier versions.