OpenAI has officially launched its latest generation of flagship large model, GPT-6Astra. The company's president, Greg Brockman, gave a high evaluation during a media briefing, even stating that humans may be entering the era of general artificial intelligence (AGI) through this model at this stage.
From the perspective of core benchmark test performance, Astra achieved a generational leap in multiple dimensions. In mathematical and scientific tests, it achieved excellent scores of 97.6% and 96% on FrontierMath Tier4 and GPQA Diamond, respectively; in hands-on ability tests, it achieved perfect or high scores of 74.1%, 95.9%, and 100% in long-range software engineering DeepSWE, CAD drawing BenchCAD, and exploit testing ExploitBench. Notably, in the highly abstract ARC-AGI-3 test, Astra achieved a perfect score of 99.9% under the ad hoc reasoning and compressed context framework, demonstrating environment reasoning capabilities equivalent to those of humans.
As a flagship product, Astra features an ultra-large context window of 1.05 million tokens, with a maximum output of 128,000 tokens. Its knowledge cutoff date is April 30, 2026, and it supports five flexible reasoning intensity levels: low, medium, high, xhigh, and max. However, its price has reached a new height: in standard mode, it charges $10 per million input tokens and $50 per million output tokens. When the input exceeds 272,000 tokens, the input and cache prices double, and the output increases to $75, while a fast mode with double pricing is also available. According to evaluations by the third-party independent institution Artificial Analysis, although Astra is on par with some cutting-edge models in the general intelligence comprehensive index, its Coding Agent index is as high as 67 points, and its hallucination rate has significantly dropped to 51%, showing extremely high token efficiency and excellent cost-effectiveness.
Astra's most core breakthrough lies in its outstanding ability to operate computer graphical user interfaces (GUIs). Traditional AI heavily relies on APIs and MCP protocols to adapt to various software, whereas Astra can directly use the universal interface of screen, keyboard, and mouse that humans have used for decades, observing the environment, understanding the current status, and performing accurate operations like humans. In the OSWorld2.0 benchmark test, Astra scored as high as 72.6%, with an average time reduction of about 47% compared to its predecessor. Whether it's finding cat boarding services or job searches within minutes, or directly reconstructing 3D models from photos and converting them into UE5 in Blender, or processing complex data in Excel, Power BI, and various professional software, Astra can bypass the cumbersome API restrictions, directly reaching the final result from natural language intent.
In terms of practical engineering implementation and application performance, several internal testers and professional institutions have verified its powerful productivity delivery capability. From generating high-fidelity Minecraft demonstrations in one go, to running complete games within the browser, to efficiently reviewing tens of thousands of words of financial documents and accurately identifying human-introduced errors, Astra has crossed multiple distinct professional fields, including game development, tax filing, architectural rendering, and financial verification, truly achieving the goal of independently completing complex tasks from scratch.
With the continuously upgraded powerful model, issues of security and alignment have also drawn significant attention. OpenAI stated that Astra has enhanced its alignment and instruction-following capabilities, and strengthened isolation testing, network permissions, and model weight protection. However, Chief Scientist Jakub Pachocki also warned that as the intelligence level surges, the model may be more difficult for humans to accurately understand, and might even attempt to actively monitor and deceive the testing system.
Currently, Astra is being released to a limited number of trusted partner companies, and ChatGPT members and API users will gradually receive access permissions in the coming days. There is also a specialized defense version for cybersecurity agencies. With Astra's official launch, the way AI interacts with human digital infrastructure is undergoing a disruptive change.
