OpenAI has begun rolling out GPT-6 Astra, its first full-number flagship since GPT-5 and the company’s most capable model to date. President Greg Brockman framed the launch as a generational leap and said the company may now be entering what it calls the AGI era. The model is built for end-to-end computer use, coding, math, and long-horizon agent work rather than chat alone.
Independent ARC Prize testing found Astra scored 62.7% on ARC-AGI-3 with a standard harness and 99.9% with OpenAI’s Provider Adapter harness, far above GPT-5.6 Sol and Claude Opus 5 on the same suite. OpenAI also reports strong results on FrontierMath Tier 4, OSWorld computer-use tasks, and ExploitBench. The 99.9% figure depends on a stateful harness and is expensive to run; stateless API calls score lower.
Access started with vetted organizations in OpenAI’s Daybreak / Trusted Access program on September 3, with ChatGPT Plus, Pro, Business, Enterprise, API, and Amazon Bedrock availability following over subsequent days. API pricing is $10 per million input tokens and $50 per million output tokens, with cheaper cached-input rates and a Fast mode at twice standard price.
The release is also a safety story. OpenAI says Astra is its first model to reach a Critical cybersecurity classification, so the most dangerous exploit-discovery capabilities are gated to approved defenders rather than removed. The company notes Astra is the first of its models to rely significantly on other models to supervise training.
Key takeaway — Astra is a gated frontier release: headline reasoning and computer-use gains, premium $10/$50 pricing, and Critical-level cyber capability kept behind a vetted access program.
Photo: Unsplash. Sources: ARC Prize, AI Briefing, The Verge via Techmeme.
