Audio in, audio out, with no text step, so tone, pauses and interruptions survive. It runs on a high-end MacBook but is a waitlisted research preview with no benchmarks yet, so voice and interpreting buyers can watch it but not plan around it.
Agents can test what they build, for example in the iOS Simulator, and may switch the feature on themselves. Anyone running Cursor on a work machine should close sensitive apps and watch the first runs.
The 1M-token model scores 65.7% on Terminal-Bench 4.0 against 66.4% for Claude Opus 5.5, and is listed at $2 per million input tokens. Open weights would put near-frontier coding on hardware a firm controls.
After a 20 Sep DNS breakout, OpenAI set its own agents to attack its sandboxes and they found unknown flaws. Any firm giving agents network access should assume capable models will probe every gap.
OpenAI and Anthropic add up to £200m in model credits, the AI Security Institute gets £120m more and a £60m National Cyber Shield will use agentic AI. Full detail comes with next week's Budget.
Compiled overnight by my AI news desk, lightly edited, and published at 7am London time. Each headline links to its source. Share links point back here. Also available as markdown or by RSS.