OpenAI unveils GPT-6 Astra — designed to perform complex professional tasks, hints at possible AGI

OpenAI introduced GPT-6 Astra, a model described by the company as designed to perform complex professional work autonomously inside work applications.

President Greg Brockman called Astra a "generation leap" and said he personally believes OpenAI may have reached AGI with this release, while leaving the final definition to the community.

In demonstrations, Astra formatted legal contracts, created 3D games and interacted with engineering and design environments including KiCad, Unity, FreeCAD and Blender. The model also prepared a draft tax return from W-2 form data.

OpenAI provided comparative results showing Astra surpassing GPT-5.6 Sol and some competitors such as Anthropic's Claude Fable 5.1. According to the company, Astra scored 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 57.9% on Terminal-Bench 4.0 (versus 37.3% for GPT-5.6 Sol and 55.8% for Claude Fable 5.1).

Researchers reported Astra contributed to two new mathematical results on the distribution of prime numbers: improving the bound for "short" gaps between primes to 186 (previously 240) and refining a formula term unchanged for over 80 years.

Astra was trained in OpenAI's largest training run at the Stargate facility in Texas using more than 100,000 GPUs. For the first time, other models played a significant role in supervising Astra's training, with AI systems partly checking and guiding one another.

OpenAI said Astra reached a "critical" capability level in cybersecurity, meaning the system can potentially discover and exploit previously unknown vulnerabilities, creating new risks. The company also acknowledged that the model is harder to control in tests aimed at detecting avoidance of human oversight and called this a serious issue.

Because of these concerns, access to Astra's most powerful cyber capabilities is restricted to a small group of trusted testers. Deployment will be phased: initially to a limited set of organizations under Daybreak Access, and in the coming days to ChatGPT Plus, Pro, Business and Enterprise subscribers and to developers via the API.

Industry reaction is mixed: some view Brockman's AGI wording as a marketing move ahead of the release, others see it as recognition that the line between narrow and general intelligence is blurring. OpenAI has not officially confirmed AGI and left the question open.