OpenAI's strongest model launches to 20 partners, DeepSeek open-sources fast inference, humanoid robots get safety stack

This week's AI news spans the frontier, the physical edge, and real-world deployment. OpenAI released GPT-5.6, its strongest model to date, under three tiers: Sol (flagship with Ultra Subagent Mode and Max Reasoning), Terra (matches GPT-5.5 at half price), and Luna (lowest cost). However, due to the June 2 executive order, initial API and Codex access is limited to roughly 20 vetted partners, with broader availability 'coming soon.' DeepSeek open-sourced DeepSpec, an MIT-licensed codebase for training and evaluating speculative-decoding draft models (DSpark, DFlash, Eagle3), making the recipe for faster inference public. A new paper, InfoKV, shows that discarding 87% of an LLM's memory can improve answers by keeping only tokens via predictive entropy and layer-wise change, beating the full-cache baseline on long contexts up to 64k tokens.

On the physical side, NVIDIA launched Halos for Robotics, the first full-stack safety system for humanoid robots, including IGX Thor compute, Holoscan sensor bridge, and a Halos OS safety layer. Agility will build it into Digit, already deployed in Amazon warehouses. China's SAMR issued the first national standard for AI-agent interconnection, defining unified identifiers for secure cross-domain agent interaction. In applied AI, GPT-5 Pro cracked a three-year immunology mystery at The Jackson Laboratory by proposing a mechanism for T cell behavior that matched a held-out experiment, described as 'a remarkable insight.' Quantifind raised $200M from Summit Partners, Citi Ventures, and S&P Global to run governed AI agents against financial-crime alerts, serving six of the top ten banks.