OpenAI Launches GPT-6 Astra With 99.9% on ARC-AGI-3 and 1M-Token Context Window

My take: OpenAI introduced GPT-6 Astra on September 3 as "the most intelligent and aligned model in the world," with benchmarks that demand attention: 99.9% on ARC-AGI-3, 88% on SRE-Bench, and 72.6% on OSWorld 2.0. Before treating those numbers as definitive, something needs to be said clearly: OpenAI tested its own model. When the company that builds the product also runs the benchmarks, we can never know with certainty whether the methodology is free of bias. These figures are worth reading with your own judgment and waiting for independent third-party evaluations.

What is concrete and actionable is what the model enables: a one-million-token context window, real gains in long-running agentic tasks, advanced coding, and autonomous computer use. For those building AI-powered workflows, that represents a genuine leap in what can be automated and delegated without constant intervention.

What tasks in your work could benefit from an agent that runs autonomously for hours on large documents or complex codebases?