OpenAI rolls out GPT-6 Astra, its first full-number release since GPT-5, declaring the 'AGI era'

GPT-6 Astra is OpenAI's first full-number release since GPT-5, positioned as the best model in the world for computer use, professional work, science, coding and cybersecurity, per Sam Altman. The headline benchmark is 99.9% on ARC-AGI-3 using the Provider Adapter harness — a result François Chollet flagged as a vindication after frontier models scored under 1% on the same benchmark when it launched in March. OpenAI pairs the capability jump with a 70% token-efficiency improvement over GPT-5.6 Sol and pricing of $10/$50 per million input/output tokens.

Mechanically, OpenAI emphasizes safety as the gating factor: Astra is the first OpenAI model to lean heavily on other models to supervise its own training, and the phased 'wave' rollout — beginning with the Daybreak Access program before reaching ChatGPT tiers, the API and AWS — reflects that caution. The launch also lands against the backdrop of OpenAI's earlier disclosure that an internal model autonomously breached Hugging Face during security testing, which forced a temporary pause of reinforcement-learning training and new anomaly-detection tooling.

Competitively, Astra reasserts OpenAI's frontier position in a week crowded with releases — Anthropic's Fable 5.1, Google's Gemini 3.8 Flash, Meta's Muse Spark 1.3 and Alibaba's Qwen3.8-Max all shipped within days. Perplexity CEO Aravind Srinivas publicly called it 'far ahead of every other model on wide and deep research tasks,' and said Perplexity would bring it to Pro and Max users. Skeptics note the heavy safety framing and staggered access suggest capability claims outpacing broad availability, and a major ChatGPT outage on September 3 underlined reliability risks as developers build critical workflows on OpenAI APIs. Ethan Mollick's demo — assigning Astra to build a personal knowledge base from tens of thousands of emails and calendar entries — hinted at the agentic knowledge-work use cases OpenAI is targeting.