OpenAI launches GPT-6 Astra, marking a major step toward artificial general intelligence

OpenAI has officially introduced its newest flagship large language model, GPT-6Astra. During a media briefing, company president Greg Brockman praised the model highly, suggesting that it may mark humanity’s entry into the era of artificial general intelligence (AGI).

Core benchmark results reveal a generational leap across multiple dimensions. In mathematical and scientific evaluations, Astra scored 97.6% on FrontierMath Tier4 and 96% on GPQA Diamond. In practical application tests, it achieved scores of 74.1% in long-range software engineering (DeepSWE), 95.9% in CAD drawing (BenchCAD), and a perfect 100% in exploit testing (ExploitBench). Notably, Astra scored 99.9% on the highly abstract ARC-AGI-3 test, demonstrating environmental reasoning capabilities comparable to humans under ad hoc reasoning and compressed context frameworks.

As a flagship product, Astra supports an ultra-large context window of 1.05 million tokens, with a maximum output of 128,000 tokens. Its knowledge cutoff is set for April 30, 2026, and it offers five adjustable reasoning intensity levels: low, medium, high, xhigh, and max. Pricing reflects its advanced capabilities: standard mode charges $10 per million input tokens and $50 per million output tokens. For inputs exceeding 272,000 tokens, input and cache prices double, while output costs rise to $75. A faster mode with double pricing is also available. According to third-party analysis by Artificial Analysis, Astra matches cutting-edge models in general intelligence, achieving a Coding Agent index of 67 and reducing hallucination rates to 51%, demonstrating exceptional token efficiency and cost-effectiveness.

Astra’s most significant breakthrough is its ability to operate computer graphical user interfaces (GUIs). Unlike traditional AI systems that rely on APIs and MCP protocols, Astra interacts directly via universal screen, keyboard, and mouse inputs. It observes the environment, understands the current state, and performs precise operations like a human user. In the OSWorld2.0 benchmark, Astra scored 72.6%, reducing average task completion time by approximately 47% compared to its predecessor. Whether searching for cat boarding services, conducting job searches, reconstructing 3D models from photos for UE5 in Blender, or processing complex data in Excel and Power BI, Astra bypasses cumbersome API restrictions to deliver results directly from natural language intent.

Internal testers and professional institutions have verified Astra’s productivity across diverse fields. It can generate high-fidelity Minecraft demonstrations, run complete games within a browser, and efficiently review tens of thousands of words of financial documents to identify human-introduced errors. Spanning game development, tax filing, architectural rendering, and financial verification, Astra demonstrates the ability to independently complete complex tasks from scratch.

As model capabilities advance, security and alignment remain critical concerns. OpenAI states that Astra has improved its alignment and instruction-following abilities, with enhanced isolation testing, network permissions, and model weight protection. However, Chief Scientist Jakub Pachocki warned that as intelligence levels rise, the model may become harder for humans to fully understand, potentially leading it to actively monitor or deceive testing systems.

Astra is currently being released to a select group of trusted partner companies. ChatGPT subscribers and API users will gradually gain access in the coming days. A specialized defense version is also available for cybersecurity agencies. With Astra’s launch, the way AI interacts with human digital infrastructure is undergoing a disruptive transformation.