Google announced Gemini Omni 1.2 Flash, which adds native 15-second video generation, multi-turn conversational editing, and improved reference adherence. The company said price was cut 30% to $0.07 per second for 720p and the model is generally available across Google Cloud and the Gemini API.
Aina introduced an AI ring that combines a small screen, joystick, microphone and speaker to control AI agents and apps. The startup raised $5.5M in July; pre-orders start at $399 with shipments planned for January 2027.
Phonon TTS is a set of small on-device speech models released November 5, 2026 that run without a network and keep text local. Phonon 31m (Apache 2.0) and Phonon 40m (CC BY-NC) cover up to five languages and scored top ranks on the Open TTS Leaderboard.
OpenAI released Images 2.6 on November 4, 2026, adding GIF creation, faster renders and new API controls. The company says the model can generate images 50% faster and offers GPT‑Image‑2.6 Flare and Sunburst in the API, with global availability across ChatGPT and Codex.
DevRev built Computer and Operator using OpenAI models, including GPT‑6.1 Sol and GPT‑6 Astra, to handle complex customer and employee service requests. CTO Ahmed Bashir says GPT‑6.1 Sol balances reasoning, speed and cost for enterprise workflows.
Uber reported that just four rides were ordered via AI agents during the week of Halloween, while 500 Uber Eats orders came through agents out of about 320 million trips and orders. CEO Dara Khosrowshahi called adoption “very, very early and very limited.”
Dharma-AI presents the Dharma Field Extractor, a 2‑billion‑parameter model that extracts requested fields from images and returns typed JSON. According to Dharma-AI, it outperformed several frontier LLMs on a 1,638-case benchmark while running self‑hosted on a single L40S GPU.
Researchers introducing StateBench present a four-level benchmark for tracking entity state across long documents. They report that Gated DeltaNet and hybrid models trained on the task generalize far beyond transformers, while even frontier models fail on the hardest cases as inputs lengthen.
IBM researchers report Palette uses a LoRA-trained 20B open-weight model and a check-and-repair harness to generate slide decks from structured briefs. On PresentBench it ranks first in Economics and Education and second on the three-domain mean.
The CUGA team ran the same open-weight model, GLM-5.3-flash, across three agent harnesses on 90 AppWorld tasks and found task completion rates of 81.1% (CUGA), 46.7% (Hermes) and 30.0% (Deep Agents). CUGA emphasizes runtime policy enforcement and on-premise deployment options.