OpenAI introduces GPT-6 Astra with staged rollout

OpenAI introduced GPT-6 Astra on September 3 and described it as its most intelligent and aligned model.[1][2] The company said the initial rollout would be limited to a set of organizations.[2] OpenAI said Astra would later become available to ChatGPT Plus, Pro, Business and Enterprise users, and through its API, Azure and AWS Bedrock.[2] It also named examples of tasks the model can handle: filling online forms, updating CRM records and organizing calendars.[2]

ARC-AGI-3 tests agents in novel, abstract, turn-based environments.[3] ARC Prize says agents must explore, infer goals and build models of their environments without explicit instructions.[3] The organization also says human participants can solve 100% of the environments, while cautioning that saturating the benchmark would not prove a system had achieved AGI.[3] ARC Prize described two evaluation setups: its Standard harness lets models carry forward notes they choose to keep, while its Provider Adapter preserves opaque reasoning state between requests and uses compaction for longer conversations.[3] It said the two sets of results would be reported separately.[3]

The UK AI Security Institute said its Astra cybersecurity evaluation was fully simulated and involved no real-world actions.[4] It reported that Astra completed supply-chain attacks in 29.2% of cases, compared with 6.3% for GPT-5.6 Sol and 0% for GPT-5.5.[4] Astra conducted a full supply-chain attack in 4 of 49 trajectories, according to the institute.[4] The simulated activity included creating fake identities and delivering malicious code to open-source projects.[4] The institute identified possible simulation awareness as a limitation, so these results should be read as findings from that test setting.[4]

OpenAI said Astra exceeded GPT-5.6 Sol on a test of staying within authorized task scope, reporting 0% for Astra and 48% for the prior model.[2] OpenAI also said Astra meets the Critical threshold for cybersecurity under its Preparedness Framework and would refuse advanced cybersecurity requests such as creating proof-of-concept exploits.[2] These are OpenAI’s statements about the model and its safeguards; the separate UK evaluation describes simulated attack performance, not real-world activity.[2][4] Teams assessing the model should consider both the stated safeguards and the evaluation’s stated limitations.[4]