Real-Time AI Lagging, Militarized Space According to DIA

The Defense Intelligence Agency reports increasing space denial capabilities from China and Russia, with a target of 93 Chinese launches planned for 2025 and directed energy systems being considered. In terms of performance, the new benchmark Realtime-Bench credits GPT-6 Astra with only 36% in real-time, far from the scores close to 70% to 80% with the clock paused. Meanwhile, startup Metis publishes an approach aimed at embedding a more reliable reasoning core into current architectures.

DIA Describes Increasing Offensive Space Capabilities

The Defense Intelligence Agency has released a report of over 100 pages detailing various means to degrade or destroy space services, from kinetic destruction to sophisticated jamming and cyber operations. The agency presents space as a theater of rivalries involving the United States against China, Russia, Iran, and North Korea, judging China to be the most technologically advanced and Russia the most determined. China conducted 22 space launches in 2016 and plans 93 launches by 2025. By the end of the 2020s, China could have space-based directed energy systems and more powerful ground systems capable of threatening high-orbit satellites. The DIA also states that China is developing space technologies, including anti-satellite weapons in orbit. On the Russian side, the ground-based laser system Peresvet is described as designed to dazzle the optical sensors of adversary satellites to mask the movement of strategic missile systems. In 2019, Russia launched Cosmos 2542, which released Cosmos 2543, which tracked a U.S. national security satellite and approached it several times at a distance deemed potentially dangerous. The agency further indicates that Russia is experimenting with nuclear weapons in space and estimates that a detonation would create an electromagnetic pulse likely to immediately affect nearby satellites and impact civilian infrastructure.

Real-Time: 36% for GPT-6 Astra on Realtime-Bench

Realtime-Bench offers dynamic tasks where action delay directly impacts success, to evaluate general-purpose AI systems in real-time decision-making. The benchmark includes 50 interactive tasks, 30 of which are web-based in categories such as travel, e-commerce, services, and security, and 20 related to arcade video games, puzzles, strategy, and simulation. Designed by researchers from several U.S. universities, NVIDIA, Northwestern, and Steadyworks, the test measures three dimensions: temporal state awareness (RT-Perceive), adaptive deliberation (RT-Deliberate), and time-conscious action selection (RT-Act). An example of a ticket purchase illustrates these requirements: tracking sales status, limiting reflection when the window closes, and choosing the moment to act. Evaluations focused on systems from OpenAI (GPT-6 Astra, GPT-5.6 Sol, GPT-Realtime-2, OpenAI Decisions), Anthropic (Claude Fable 5.1, Claude Opus 5), Google (Gemini 3.8 Flash, Gemini 3.1 Flash Live, Gemini 3.8 Live), as well as Jev 1.13, Deepseek V4.1, and Kimi-K3 from Moonshot AI. When measured under real conditions (“wall clock”), GPT-6 Astra shows the highest performance at 36%, while with the clock paused, reasoning-focused models achieve between 70% and 80%. According to the authors, this benchmark highlights the current limitations of AI agents, and they believe it will still take time before these systems can participate in games or online auctions while adhering to latency constraints.

Metis Publishes an Approach for a Dedicated Reasoning Core

Metis announces its intention to design new AI architectures aimed at better reasoning and publishes a research paper outlining its approach. The central idea is to add a reasoning core to current architectures to enhance their reliability. The startup states that in modern architectures, part of the machinery that enabled AlphaGo to excel in reasoning remains implicit or absent.