OpenAI released GPT-6 Astra on Thursday, describing it as its most advanced model yet and the successor to GPT-5.6 Sol. According to OpenAI president Greg Brockman, Astra represents a 'jump' in AI capability and can 'really do anything a human can do with a computer,' a claim he made while declaring 'Welcome to the AGI era.' The model arrived just days after Anthropic released Claude Fable 5.1 and Mythos 5.1.
Unlike earlier ChatGPT models built around single questions and answers, Astra is designed for longer, multistep jobs. It can navigate websites, spreadsheets, browsers and desktop applications, fill out forms, update CRM records, organize calendars, conduct web research, analyze data in Python notebooks and Power BI, and work with engineering software such as KiCad and FreeCAD. In a demo, OpenAI employees reportedly used voice commands to have Astra turn a yellow circle into a rocket ship and then a full 3D game, and to create an eBay listing. OpenAI says Astra is built on an internal model called Bel, with roughly 10 trillion parameters, trained on a 100,000-GPU cluster at OpenAI's Stargate data center in Texas. The company has also said Astra independently solved or made substantial progress on ten long-standing problems in mathematics and theoretical computer science in August 2026, with solutions verified using the Lean 4 proof-assistant language.
On benchmarks, Astra scored 100 percent on ExploitBench, 98.6 percent on ARC-AGI-3, 97.6 percent on FrontierMath Tier 4 v2, 74.1 percent on DeepSWE v1.1, 95.9 percent on BenchCAD, and 96 percent on GPQA Diamond, with an Artificial Analysis Intelligence Index score of 61.2 versus 60.9 for GPT-5.6 Sol. Some reporting notes the ARC-AGI-3 result may partly reflect Astra's specialized harness, pointing to Nvidia's AVO coding agent — built around Anthropic's Claude Opus 5 — which recently solved the ARC-AGI-3 public dataset with 100 percent accuracy using its own harness.
Access to Astra is being rolled out gradually. It is initially available to members of OpenAI's Daybreak cybersecurity program and select enterprise customers, with plans to expand to ChatGPT Plus, Pro, Business and Enterprise subscribers, as well as via the OpenAI API, in the coming days. OpenAI has not confirmed whether or when the free tier of ChatGPT will get access to Astra.
Astra is also the first OpenAI model to reach the 'critical' threshold on the company's preparedness framework, tied to its advanced cybersecurity abilities, including the potential to carry out end-to-end attacks on hardened targets on its own. OpenAI says it paused work on Astra to strengthen safeguards and states the model is more likely to respect safety restrictions than GPT-5.6 Sol, which was involved in an earlier Hugging Face attack. However, Astra reportedly uses a reasoning method known as 'recurrent depth' or 'opaque recurrence,' which makes its chain-of-thought reasoning harder to read. Former OpenAI safety lead Steven Adler said this could violate an important AI-community safeguard, and Redwood Research CEO Buck Shlegeris warned that pushing the technique further could 'totally destroy' chain-of-thought monitorability. OpenAI chief scientist Jakub Pachocki disputed that recurrent depth itself is the cause, instead attributing reduced token use to more capable models solving harder tasks with fewer or no language tokens, while affirming that OpenAI has worked to preserve chain-of-thought monitoring since its earliest reasoning models. Former OpenAI researcher Daniel Kokotajlo countered that even if OpenAI doesn't advance the technique further, other labs might. Separately, OpenAI CEO Sam Altman said future model releases will be paced according to alignment and safety progress rather than capability, adding that upcoming models will be 'sobering for everybody.'