GPT-5-based agent frameworks reportedly score 70% on OSWorld, close to a 72% human mark

One post claims GPT-5-based agentic frameworks reached 70% on OSWorld, a cross-OS computer-use benchmark, close to what it calls a 72% human mark. The author argues that human-level computer use may be within reach. Separately, another post, citing unspecified reports, says OpenAI is preparing Agent Builder, a visual workflow canvas with drag-and-drop design, MCP connectors and built-in guardrails. It expects an announcement at DevDay.