Scale AI’s defense pivot and Donovan LLM deployment

Scale Donovan and the defense pivot

Scale Donovan functions as the first large language model deployed to a U.S. government classified network. The platform ingests thousands of orders, situation reports, intelligence reports, and sensor logs to assist operators, analysts, decision-makers, and commanders. The XVIII Airborne Corps and the Marine Corps University School of Advanced Warfighting use the system to plan and act. Scale AI signed a five-year, $100 million enterprise agreement with the Department of Defense Chief Digital and Artificial Intelligence Office on September 17. This agreement follows two and a half years of prototype agreements between Scale and the agency. The contract provides access to advanced AI capabilities on networks up to Top Secret/SCI. The program seeks to achieve AI overmatch through the deployment of AI agents. The Defense Innovation Unit leads the Thunderforge program to develop AI agents for military planning and operations through a team of global technology partners including Scale AI, Anduril, and Microsoft for the Department of Defense. This program focuses on modeling, simulation, decision-making support, and automated workflows. The rollout begins with the U.S. Indo-Pacific Command and the U.S. European Command. Doug Beck, the DIU director, stated that there is a fundamental mismatch between the speed of modern warfare and the ability to respond.

The frontier AI landscape

The Department of Defense allocates $800 million to frontier AI companies through the Chief Digital and Artificial Intelligence Office. This funding distributes $200 million contracts to Anthropic, Google, OpenAI, and xAI. The program aims to embed AI in defense across warfighting, intelligence, logistics, and enterprise functions. Anthropic and Palantir partnered with Amazon Web Services to provide intelligence and defense agencies access to Claude 3 and 3.5 models. OpenAI and Anduril formed a partnership to deploy AI systems for national security missions. These systems focus on counter-unmanned aircraft systems and the ability to detect and respond to aerial threats in real-time. xAI offers Grok for Government through the General Services Administration to provide models to federal government departments and agencies. Palantir holds a five-year contract worth up to $100 million to expand military access to its Maven AI warfare program. The Pentagon FY2026 budget includes $13.4 billion for AI-facilitated autonomous systems. This investment follows increased global military expenditure, which reached $2.887 trillion in 2025. Germany increased expenditure by 24% to $114 billion, while Spain increased spending by 50% to $40.2 billion. In Asia, China marked its 31st consecutive year of spending increases, reaching $336 billion. Japan reached 1.4% of its GDP in defense spending, which is the highest since 1958. Taiwan increased spending by 14% to $18.2 billion. OpenAI received a $200 million contract earlier this year for work with Anduril to develop AI systems for national security. Boeing’s Q3 2026 revenue included 37.9% from the U.S. government, including Foreign Military Sales. RTX generated 51.1% of its Q3 2026 sales from U.S. and foreign government customers. Lockheed Martin generated 69.0% of its Q3 2026 sales directly from the U.S. government. Booz Allen Hamilton derives approximately 98% of its revenue from U.S. government contracts. You already know that defense procurement cycles often lag behind commercial innovation.

Intelligence analysis model benchmarks

Model performance dictates the effectiveness of intelligence analysis. Gemini 3.1 Pro leads in reasoning with a 44.4% score on Humanity’s Last Exam and a 77.1% score on ARC-AGI-2. Gemini 3.1 Pro also holds a 94.3% score on GPQA Diamond. Gemini 3.1 Pro leads in competitive programming with an Elo of 2887 on LiveCodeBench Pro and a 68.5% score on Terminal-Bench 2.0. Gemini 3.1 Pro also leads in scientific computing with a 59% score on SciCode. Claude Sonnet 4.6 leads in agentic automation with a GDPval-AA Elo of 1633. Claude Sonnet 4.6 achieves a 61.3% score on MCP Atlas. Claude Opus 4.6 leads in computer use scenarios with a 66.3% score on OSWorld-Verified. GPT-5.2 achieves a 38.2% score on OSWorld-Verified. Claude Opus 4.6 achieves an 80.8% score on SWE-Bench Verified. Gemini 3.1 Pro achieves an 80.6% score on SWE-Bench Verified. Claude Opus 4.6 leads in web browsing with an 84.0% score on BrowseComp. Gemini 3.1 Pro achieves an 85.9% score on BrowseComp. Claude Opus 4.6 leads in multilingual coding with a 77.5% score on SWE-bench Multilingual. Gemini 3.1 Pro leads in multimodal understanding with an 80.5% score on MMMU-Pro. GPT-5.2 follows with a 79.5% score. Claude Opus 4.6 provides 32,000 max output tokens. Gemini 3.1 Pro provides 6,000 max output tokens. Claude Opus 4.6 has a 200,000 token context window. Gemini 3.1 Pro has a 1,000,000 token context window. GPT-5.2 has a 400,000 token context window. Claude Opus 4.6 costs $15 per million input tokens and $75 per million output tokens. Gemini 3.1 Pro costs $2 per million input tokens and $12 per million output tokens. Claude Sonnet 4.6 costs $3 per million input tokens and $15 per million output tokens. GPT-5.2 costs $1.75 per million input tokens and $14 per million output tokens. How will the integration of autonomous agentic workflows change the chain of command in contested environments?