The Sequence Radar #885: Last Week in AI: Models, Games, and the Future of Evaluation

TheSequence · 2 weeks ago

OpenAI released GPT-5.6 as a three-tier model suite (Sol, Terra, Luna) with structured safety architecture and phased access strategy, while Anthropic introduced Claude Tag for structured prompt interaction and General Intuition raised $320M at $2.3B valuation to train large action models on gameplay clips. The most concrete development was the LayerLens Stratix Cup, where Claude Opus 4.8 defeated GPT-5.5 1-0 in a soccer-based evaluation arena, demonstrating models executing complex autonomous behavior under real-time constraints rather than answering static questions. These releases signal a shift from evaluating models on benchmark leaderboards toward testing them in dynamic environments where they must sense, plan, and adapt, reshaping how AI development prioritizes embodied reasoning and real-world deployment safeguards.