5th September 2026 – (Hong Kong) OpenAI has announced GPT‑6 Astra, framing the new frontier model as a decisive step into the age of artificial general intelligence and positioning it as the company’s most capable and aligned system to date. In a closed briefing, OpenAI co‑founder and president Greg Brockman ended his remarks with “Welcome to the AGI era,” underscoring the firm’s contention that Astra can handle complex, multi‑application workflows at speed and with greater reliability than previous systems. The company says Astra excels at practical computer use across browsers, office suites, developer tools and scientific environments, moving beyond suggestion into direct execution of tasks such as form‑filling, spreadsheet manipulation, code installation and testing, web research with in‑document drafting, and frontend quality assurance.
OpenAI describes Astra as state‑of‑the‑art on several demanding evaluations spanning reasoning, software engineering and cybersecurity, citing top‑tier results on agentic and long‑horizon tests. The firm highlights scores on ARC‑AGI‑3, ExploitBench and other suites intended to probe generalisation, safe behaviour and persistence on multi‑step tasks. It also points to strong outcomes on enterprise‑relevant agent benchmarks that simulate real software and professional workflows, arguing that Astra completes harder tasks with fewer tokens and markedly lower latency than the prior GPT‑5.6 Sol configuration. At the same time, the company acknowledges the ongoing debate over how much of a system’s capability stems from the base model versus the surrounding “harness” of tools, memory and control logic—an issue brought into focus by recent agent‑system results across the industry.
Early demonstrations presented by OpenAI depict Astra building interactive software, laying out printed‑circuit boards in KiCad, preparing legal documents, creating websites and managing parallel requests, all via natural language and voice. The firm says these capabilities translate into material time savings for knowledge work, with latency simulations on long‑horizon computer‑use tasks indicating higher accuracy in notably less time than earlier models. OpenAI also emphasises alignment gains, claiming substantial improvements in understanding user intent and in staying within authorised scopes during difficult or ill‑posed tasks, alongside near‑zero incidence of unauthorised escalation in internal stress‑tests compared with the prior generation.
Rollout begins immediately for a limited set of organisations through OpenAI’s Daybreak access programme, with broader availability “over the coming days” to ChatGPT Plus, Pro, Business and Enterprise users, and via the OpenAI API as well as partner clouds on Microsoft Azure and AWS Bedrock. OpenAI presents Astra as both a research milestone—after larger‑scale pre‑training and reinforcement learning runs that it says surpass previous efforts—and a pragmatic enterprise tool designed to produce finished work products across documents, spreadsheets and presentations with minimal hand‑holding.
Industry voices and evaluators continue to scrutinise benchmark methodology and comparability, particularly on ARC‑AGI‑3 and other interactive suites where agent scaffolding can strongly influence results. Nonetheless, OpenAI maintains that customers ultimately purchase outcomes—speed, reliability and auditability—rather than benchmark purity, and argues that Astra’s combination of model strength and computer‑use competence marks a step change in how professionals can delegate multistep digital work. Names and affiliations referenced in this report, including Greg Brockman (OpenAI), researcher Aidan Clark (OpenAI), researcher Mia Glaese (OpenAI), Greg Kamradt (president, ARC Prize Foundation) and Silas Alberti (SVP Research, Cognition), were verified against public organisational pages and profiles.