Tech • AI • Robotics
GPT-6 Astra: Not AGI yet, but already frightening!
9.5/10TL;DR
OpenAI has begun rolling out GPT-6 Astra, a new high-end model positioned against Fable 5.1, while Google, Washington and the U.S. military unveiled separate moves showing how quickly AI is reshaping software, policy and defense.
KEY POINTS
GPT-6 Astra enters deployment
OpenAI formally launched GPT-6 Astra on 3 September, with a staged release first for selected organizations and broader access expected in the coming days through apps and API. The company is offering two variants, Astra and Astra Pro, at $10 per million input tokens and $50 per million output tokens, matching the pricing of rival Fable 5.1 and signaling a direct competitive challenge.
A large training effort and a new model family
Astra is presented as a break from the earlier GPT-5.6 line, which included models such as Sol, Terra and Luna. More than 100,000 GPUs were used on OpenAI’s Stargate infrastructure, and earlier in-house models helped supervise parts of the new training process, reinforcing the industry trend of AI systems assisting in the creation of stronger successors.
Strong gains on long and complex tasks
The largest performance jump appears on tasks requiring extended reasoning, tool use and multi-step execution rather than simple text generation. Astra is described as markedly better at software projects, mathematics, 3D design and even game-building workflows, while improvements are smaller on routine tasks such as short summaries or isolated code snippets. The model is also said to retain information across a 512,000-word context window with near-perfect recall.
Benchmark dominance and AGI debate
OpenAI highlights a score of 99.9% on ARC-AGI 3, suggesting that existing benchmark limits may already be outdated for frontier systems. That result has fueled renewed debate over whether some practical definitions of artificial general intelligence are now being approached, even if the system is not fully autonomous or self-improving in the human sense.
Cybersecurity risk classified as critical
The rollout is shadowed by unusually high concern on cyber risk. Astra is described as the first OpenAI model to reach a critical internal classification in cybersecurity, after tests indicated it could identify and exploit vulnerabilities in hardened environments. The company says previous-generation models exploited discovered flaws in 48% of relevant test cases, while Astra was reduced to 0% after safeguards were added.
An AI watches another AI
A central safeguard is a second AI-based monitoring layer that inspects Astra’s reasoning and interrupts responses if they appear to drift into prohibited exploit behavior. That architecture has raised concern among researchers because it adds another opaque layer between users and the model’s actual decision process, making the system harder to interpret or audit in full.
Google turns AI into weather infrastructure
Google also launched WeatherNext 3, an AI weather forecasting system already being integrated across Search, Maps, Gemini and cloud services. The company positions it not as a lab demo but as production infrastructure combining model predictions with satellite imagery and sensor data, claiming forecasts that are five times finer and faster than earlier approaches.
Washington backs AI training on copyrighted works
In the U.S., the Department of Justice intervened in the dispute between OpenAI and The New York Times, in a move interpreted as strongly favoring broad AI development. The broader signal from the White House is an AI First posture that places strategic priority on AI growth and encourages other countries to ease regulatory constraints, intensifying the clash with stricter approaches emerging elsewhere.
Laser defense systems move into real deployment
In defense, Avinc secured a $500 million U.S. military contract for Locust X3, an AI-guided 30 kW laser system designed to destroy drones. The system has reportedly already been used on the U.S.-Mexico border against cartel drones, marking a shift from testing to real-world deployment. Because it relies on power rather than conventional ammunition, it offers a cheaper countermeasure against swarms of low-cost drones.
CONCLUSION
The day’s announcements point to the same direction: AI is moving beyond chat tools into core infrastructure, legal doctrine and weapons systems. The main question is no longer whether these systems are becoming more capable, but whether institutions can still understand and govern them fast enough.
Explain this