Particle.news

OpenAI Releases GPT-6 Astra, a Model That Can Act Inside Software

OpenAI says Astra’s ability to complete multi-step tasks inside apps triggered safety gating and points to a move toward premium, autonomous workflows.

Overview

  • OpenAI released GPT-6 Astra on Sunday as a frontier model built to operate inside browsers, spreadsheets, desktop apps and design tools to complete multi-step tasks rather than only giving instructions.
  • Astra is designed to produce finished outputs and keep memory across long sessions, letting it build slide decks, code for extended debugging sessions, and turn 3D assets into game-ready scenes with less human hand-holding.
  • OpenAI labeled Astra 'Critical' under its Preparedness Framework and is gating the model’s offensive exploit-creation features after reporting 100% on its ExploitBench and strong reverse-engineering results for defensive use.
  • Those security and performance claims are self-reported and come with caveats because independent tests and different evaluation setups show much lower scores on some benchmarks.
  • OpenAI is rolling Astra out in stages to vetted Daybreak partners and paid users through the API, Microsoft Azure and Amazon Bedrock with premium pricing aimed at high-value agentic tasks, and the launch has drawn sharp criticism from creative communities over advertising and potential content misuse.