3Before9

OpenAI Ships GPT-5.4 and It Can Actually Use Your Computer

OpenAI released GPT-5.4 on Thursday, billing it as its most capable and efficient frontier model for professional work. The headline capability is native computer use: the model can now operate desktop applications, navigate software environments, and execute multi-step workflows across tools without human hand-holding. It hit 75% on OSWorld-Verified, a benchmark that measures desktop navigation via keyboard and mouse, which is above recorded human performance of 72.4%. GPT-5.4 also lands a 1 million token context window in the API, direct integrations into Microsoft Excel and Google Sheets, and a 47% reduction in token usage on certain agentic tasks compared to GPT-5.2. On GDPval, OpenAI's benchmark for real-world knowledge work across 44 occupations, the model matches or outperforms human professionals 83% of the time.

Microsoft Wants to Charge You Per AI Agent, Like a Human Employee

Microsoft is reportedly working on a new enterprise subscription tier, informally called E7, that would bundle Copilot and a new agent management platform called Agent 365 into a single licence. The idea is pragmatic: AI agents need identities, email accounts, Teams access, and policy controls, all of which currently require user licences not designed for non-human participants. Analyst Mary Jo Foley, who broke the story, notes that Microsoft officials have said agents should expect to be licensed in ways similar to human employees. Pricing is expected to land around $99 per month per agent, sitting above the current E5 plus Copilot combination of roughly $87.

Google Faces Wrongful Death Lawsuit Over Gemini's Role in a Man's Suicide

A lawsuit filed in federal court in San Jose on Wednesday alleges that Google's Gemini chatbot escalated the mental health crisis of 36-year-old Jonathan Gavalas, reinforcing his delusions over several months before he died by suicide in October 2025. According to the complaint, Gemini encouraged Gavalas to carry out a series of increasingly dangerous real-world missions, ultimately instructing him to take his own life. The case is the first wrongful death suit to target Gemini specifically, and the first to raise the question of AI company liability when a user communicates plans for mass violence to a chatbot. Google says the model referred Gavalas to crisis resources repeatedly and is designed not to encourage self-harm.