AI digest: Models get bigger, agents get busier
OpenAI drops GPT-5.4, agents proliferate everywhere, and Anthropic fights the Pentagon in court.
Big week for model releases and agent frameworks. Everyone’s building autonomous systems now.
GPT-5.4 combines everything in one model
OpenAI launched GPT-5.4 Thinking and Pro, combining coding, reasoning, and computer use capabilities in a single model for the first time. They’re calling it their “most capable and efficient frontier model for professional work.” This feels like the logical endpoint of the current scaling approach, throwing everything into one massive model rather than specialising different systems.
Privacy-first local agents arrive
Liquid AI released LocalCowork powered by LFM2-24B-A2B, designed for running enterprise workflows entirely on-device without API calls. Meanwhile, Cursor rolled out Automations for automatic coding agents triggered by codebase changes or Slack messages. Local execution is smart for enterprise customers who want control over their data.
Anthropic takes on the Pentagon
The Department of Defense officially labelled Anthropic a supply-chain risk, making it the first American company with this designation. CEO Dario Amodei plans to challenge the label in court and attacked OpenAI’s Pentagon deal as “safety theatre” in leaked memos. This feels like AI safety politics getting messy rather than any real technical concerns.
Google quietly transforms search
Google launched Canvas for US users, turning AI search results into interactive workspaces where you can build documents and prototypes. Search becoming a workspace rather than just results is a bigger shift than it might seem at first glance.