AI Round-up: May 21, 2026

OpenAI supplied the day's sharpest research signal: its reasoning model disproved a long-standing discrete-geometry conjecture, while GPT-Rosalind and Ramp's Codex workflow kept the lab visible across research and developer tooling. Google kept the model-launch tail alive with Gemini 3.5 Flash and Gemini Omni posts, while Cohere open-sourced Command A+ and Perplexity pushed query-aware compression for search.

Anthropic dominated the business and policy side of the snapshot. Reuters and Reddit focused on its SpaceX compute bill, profitability claims and Mythos risk debate; Anthropic also announced a KPMG-wide Claude alliance and more frontier-AI consultation. The agent tooling lane stayed security-heavy, with Microsoft's RAMPART and Clarity, Mini Shai Hulud supply-chain warnings, OpenAI Codex on mobile and 1Password's Codex integration all pointing at production hardening rather than demos alone.