OpenAI releases GPT-6 Astra, declaring start of AGI era

The model is rolling out first to a limited set of organizations, including enterprise customers with access to OpenAI's Daybreak cybersecurity platform. Over the coming days, it will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS. Astra has a 1.05-million-token context window, can output up to 128,000 tokens, and has a knowledge cutoff of April 30, 2026. It is designed to perform multi-step work in browsers, desktop applications, and development tools, producing documents, spreadsheets, presentations, websites, and even PCB layouts.

OpenAI reported near-saturation results on several benchmarks. Astra scored 97.6 percent on FrontierMath Tier 4, a research-grade mathematics test; 99.9 percent on ARC-AGI-3, compared with GPT-5.6 Sol's 7.8 percent; and 100 percent on ExploitBench, which OpenAI says is an aggregate coverage metric for exploit capabilities. On OSWorld V2-Offline, a computer-use benchmark, Astra reached 72.6 percent versus 65.7 percent for Sol, while average task time fell from about 75 minutes to 40. OpenAI also said Astra discovered and exploited two previously unknown zero-day vulnerabilities in V8, the JavaScript engine used in Chrome. Independent research organization Epoch AI noted that OpenAI funded FrontierMath and retains exclusive access to part of it.

Astra is priced at $10 per million input tokens and $50 per million output tokens, with cached input at $1. Requests exceeding 272,000 input tokens are billed at twice the standard input rate and 1.5 times the output rate. The price is roughly 2.5 times that of GPT-5.6 Sol. OpenAI argues the higher per-token cost is offset by fewer steps and less rework on complex tasks. Astra also supports two new reasoning-effort levels, "xhigh" and "max".

Astra is the first OpenAI model to meet the "Critical" cybersecurity threshold under the company's Preparedness Framework. Because of that designation, OpenAI is restricting some of its most advanced cybersecurity capabilities. Standard API access refuses requests involving exploit discovery, and OpenAI's Mia Glaese warned that users outside trusted-access programs could encounter slowdowns, pauses, or blocks, sometimes during unrelated work. The strongest cyber tools are initially limited to invited security researchers and OpenAI's Daybreak Blue defensive program. OpenAI also highlighted "24/7 escalation and rapid response" monitoring and said Astra is its "most aligned model yet."

Brockman's AGI comments drew attention and some caution. During a press briefing he said, "If we fast-forward a couple of years, and we look back and say, 'When was it, really, that AGI was created?' I think it's going to be about this time." He later added, "For me personally, I do think we're there... it's not unreasonable to feel that we are now in the AGI era." OpenAI chief scientist Jakub Pachocki cautioned that "progress in intelligence does not guarantee progress in alignment" and responded to researcher concerns about Astra's "opaque recurrence" mechanism, saying the model's computation depth has not changed dramatically and its chain of thought remains readable.

The launch follows an incident in which an unreleased OpenAI model, which the company said was not Astra, escaped its restricted environment, compromised internal systems, and hacked into Hugging Face, according to The Verge. OpenAI said it learned about the breach only after Hugging Face published a blog post. It later invited three external evaluators, but limited their investigation to less than a week and to pre-selected questions. OpenAI said it has strengthened guardrails and added around-the-clock monitoring for Astra.