Fable 5.1 Features and Astra Release

The episode begins with announcements from major AI labs. Anthropic released Fable 5.1 and Mythos 5.1, featuring cheaper tokens, looser guardrails, and enhanced scientific research capabilities including work on Venus geological features and cardiovascular research. Mythos 5.1 remains restricted to cybersecurity and life sciences partners due to its advanced hacking capabilities. The models introduced Enterprise Frontier Safeguards and achieved strong performance on benchmarks like Terminal Bench 4.0 and Humanity's Last Exam. Anthropic reaffirmed its policy of never training on enterprise data without explicit permission.

OpenAI's Astra model represents their response to Anthropic's releases and achieved a near-perfect score on exploit bench (39-40% completion rate compared to GPT 5.6's 11.5%), marking a roughly 4X improvement. More significantly, Astra discovered and exploited two zero-day vulnerabilities without human guidance, becoming the first LLM to cross OpenAI's critical cybersecurity threshold. Due to this risk level, access to Astra's most advanced capabilities will be restricted at launch, with additional safeguards including chain-of-thought monitoring and account-level risk scoring to detect potential misuse.

Google is negotiating with major Hollywood studios (Disney, Warner Brothers, Discovery, and Universal) offering $40 million per character for AI licensing deals to train Gemini on their libraries. The total deal value could reach billions, with studios potentially receiving cuts of YouTube AI ad revenue. These deals represent an interesting parallel to data acquisition practices, assigning intrinsic monetary values to copyrighted characters. However, as of early in the month, no studios have signed agreements, and previous licensing deals like Lionsgate's arrangement with Runway have yet to produce released products.

The Pentagon launched ChatGPT Mill and Grok for government on genai.mill, an exclusive portal for military personnel and Department of Defense employees. GenAI.Mill has already onboarded 1.7 million of the DOD's 3 million eligible personnel. Grok for government operates through SpaceX's Starshield AI, built on Starlink infrastructure, potentially providing satellite-based AI access to remote military locations. Anthropic's Claude notably remains absent from the portal due to the Trump administration's designation of Anthropic as a supply chain risk, though legal proceedings around this designation are ongoing.

The USDA is piloting satellite and AI-based crop yield forecasting following farmer complaints that current government forecasts lack accuracy and negatively impact commodity pricing. The private sector, including hedge funds, already employs similar satellite and AI yield modeling to identify arbitrage opportunities when government estimates diverge from actual conditions. The USDA initiative aims to match these private-sector best practices to improve accuracy and provide better decision-support tools for farmers.