GPT-6.1 Ultrafast ⚡, Gemini universal agent 🤖, speculative decoding 🧠

OpenAI has rolled out Ultrafast for GPT‑6.1 Sol, a new API mode that delivers up to eight times the speed of Sol Standard while maintaining near‑Astra‑level intelligence. The feature is available in the API, Codex, and ChatGPT Work, with pricing set at $12 per million input tokens and $60 per million output tokens. Google has introduced a universal Gemini agent designed to autonomously handle knowledge work, media creation, coding, and multi‑step workflows across Workspace and other enterprise tools. Gemini adds persistent context, multi‑agent orchestration, model routing, governance, and cost controls. A startup called Hone, founded by former employees of fast‑growing AI firms, has secured a $60 million seed round to build AI agents that can serve as professional staffers, capable of managing long‑running tasks over weeks or months. Meanwhile, the Epoch Automation Reports indicate that current frontier and open‑weight models still struggle with both well‑defined and open‑ended tasks, lacking the judgment to fully automate complex work. Speculative decoding is presented as a technique where a small draft model proposes a likely sequence that a larger model verifies in a single forward pass, potentially speeding up generation when batch sizes are suboptimal. The ATLAS benchmark evaluates search agents on real‑world web‑search tasks, revealing that even high‑effort agents miss significant answers, underscoring room for improvement. Additional tools and projects highlighted include Quicksand, an async Python API for sandboxing AI agents in QEMU VMs; NVIDIA’s LongLive collection focused on long‑video generation; and various other agentic and security tools. These developments collectively demonstrate the industry’s focus on speed, autonomy, and evaluation of AI agents in enterprise contexts.