GPT-6 Astra: OpenAI sees AGI drawing closer. And it takes on Anthropic in the field of agents
OpenAI presents GPT-6 Astra as a generational leap towards more autonomous systems, capable of programming, using computers, analysing data and identifying problems without explicit instructions.
OpenAI has released the GPT-6 Astra model. President Greg Brockman described it as a “generational leap”, explicitly referring to the dawn of the era of general artificial intelligence.
AGI, or artificial general intelligence, has for years been a sort of horizon: everyone pointed towards it, but no one could say how far away it was. The new model promises to do something that previous chatbots could only do to a limited extent: not merely responding, but understanding an objective, using tools, identifying problems and carrying out a task with fewer human instructions. If we ask it to analyse a complex chip supply chain, Altman explains, it can go so far as to independently flag up a problem that wasn’t included in the initial request. In other words, the chatbot is starting to behave a little more like a colleague. With all the usual caveats.
What is Astra like?
OpenAI describes it as its ‘smartest and most aligned’ model. The figures serve primarily to stake out territory. Astra achieves 99.9 per cent on ARC-AGI-3, 97.6 per cent on FrontierMath Tier 4 and 100 per cent on ExploitBench, a test dedicated to cybersecurity. In the scientific tasks on Terminal-Bench, it has improved from GPT-5.6 Sol’s 22.4% to 64.6%. In coding, the improvement is from 37.3% to 57.9% on Terminal-Bench 4.0.
But the most interesting change lies elsewhere. Astra promises to be able to use a computer. It can browse the web, fill in forms, update a CRM, organise a calendar, install software, manage a website, and work on documents, spreadsheets and presentations. On OSWorld 2.0, it scores 72.6 per cent compared to its predecessor’s 65.7 per cent and, according to OpenAI, completes these tasks in around 47 per cent less time. This is where generative AI really meets the agents: less conversation, more virtual mouse.
Then there is the issue of working memory. During lengthy programming tasks, Astra can retain notes and retrieve information from previous context windows.
