OpenAI has announced GPT-6 Astra, its next-generation intelligence model. This model is said to integrate years of research in pre-training, reinforcement learning, and alignment.
Model ReleasesOpenAIGPT-6 Astra
OpenAI Announces GPT-6 Astra, Significantly Improving Computer Use and Professional Task Performance
This article is a translation. Read the Japanese original
In the field of mathematics, it recorded a score of 98% on FrontierMath Tier 4. It also achieved scores of 99.9% on ARC-AGI-3 and 100% on ExploitBench, contributing to the resolution of existing complex mathematical problems, according to the company.
Capabilities in computer use and web browsing have also reached a new level. In addition to daily tasks such as filling out online forms, updating customer records in CRMs, and organizing calendars, it is also capable of software installation and testing, as well as frontend quality assurance (QA).
In terms of efficiency, in OSWorld 2.0 simulations, it reduced the time per task by approximately 47% compared to GPT-5.6 Sol. Furthermore, when combined with updates to the Codex harness, it achieved task completion approximately 1.9 times faster on the Mind2Web benchmark.
Regarding safety, alignment has been strengthened to understand user intent and control model behavior. According to developed evaluation metrics, while 48% of cases in GPT-5.6 Sol involved operating outside the intended scope when facing difficult challenges, GPT-6 Astra recorded 0%.
Availability begins today for select organizations. Following this, it is scheduled to be rolled out to ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock.
Source: GPT-6 Astra: A new generation of intelligence (OpenAI News, 2026-09-03)