Shandong University researchers found harness choice alone swung SWE-bench cost up to 3x for the same model, with success within two points. See how Claude Code, OpenCode and mini-SWE-agent compared.

Workspace admins can filter and export users' Gemini conversations in Vault for legal and compliance review, per Google Workspace Updates. See what Vault exports and what is still missing.

Teams will score the conversation in first-contact external chats for phishing and impersonation, reaching GA in November 2026, per a Microsoft 365 message center notice. See what Teams can now catch.

Future OpenAI models will train on cached and simulated web content, per OpenAI's blog, two weeks after Anthropic cut live access from its internal evaluations. Read the agent incidents that forced the switch.