AI Round-up: May 6, 2026

xAI supplied the largest single launch signal with Grok 4.3 on its API, pitching the model around agentic tool calling, instruction following and enterprise-domain performance. Around it, Anthropic and Microsoft pushed agent infrastructure into concrete work settings through finance and insurance agents, Dataverse as an agent data platform, Dynamics case-enrichment simulation and Copilot Cowork, while OpenAI’s GPT-5.5 Instant and memory updates kept ChatGPT in the launch conversation.

The debate layer stayed focused on control and capability: Anthropic Fellows research on models holding back capability, Reuters coverage of AI model stress tests, Microsoft’s AI prediction-quality tooling and Reddit’s SB3444 safety discussion all kept evaluation and governance in frame. Developer chatter remained centred on Codex and Claude Code, including npm-download spikes, local-LLM workflows and Andrej Karpathy’s agentic engineering thread.