OpenAI has started a limited rollout of GPT-6 Astra, a model built to carry out complex, end-to-end tasks rather than simply return chat responses. The key change is its reported ability to work across browsers, spreadsheets, websites and desktop applications, allowing it to complete multi-step workflows using software tools.
Astra can reportedly generate documents and presentations, fill in forms, update CRM records and use tools including Power BI, KiCad and FreeCAD. In demonstrations, it has been shown responding to voice commands to edit objects, create a 3D game and prepare an eBay listing without the user manually directing every stage.
The model uses a multi-agent approach for broader assignments. A main agent can break a problem into smaller tasks, assign them to sub-agents and consolidate the results, enabling work on the same objective over longer periods while moving between different tools.
OpenAI’s preliminary figures place GPT-6 Astra at 100% on ExploitBench, 98.6% on ARC-AGI-3, 97.6% on FrontierMath Tier 4 v2 and 74.1% on DeepSWE v1.1. OpenAI co-founder and president Greg Brockman described the development as the start of an “AGI era.”
Those results require context. The reported ARC-AGI-3 score may be affected by the evaluation harness used in testing. An NVIDIA agent built around Claude Opus 5, for example, reportedly completed the public ARC-AGI-3 set using its own infrastructure. Near-perfect benchmark scores therefore do not by themselves establish that general artificial intelligence has been achieved.
Astra arrives less than two months after GPT-5.6 Sol, underlining the pace of OpenAI’s model releases. Access is initially restricted to a small group of customers, while Sam Altman has indicated that future launches may be governed by advances in safety and alignment rather than technical capability alone. The central unanswered question is how quickly OpenAI will expand access to increasingly autonomous agents as its control mechanisms mature.
