AI models, services and robotics
Google's real-time voice model Gemini 3.8 Live now has a face that matches lip movements and expressions. Businesses can build a brand avatar from a single reference photo and have it converse in 97 languages.
A pioneer of meta-learning and world models is joining Tokyo's recursive self-improvement lab. An idea from a 1987 diploma thesis has become one company's research roadmap 39 years later.
Liquid AI has released a roughly 280-million-parameter draft model that attaches to its vision-language model LFM2.5-VL-3B. It leaves the answers unchanged while speeding up decoding by up to 3.13 times.
Perplexity has built Fast Search, an agent-focused Search API preset, on top of its new search engine Photon. Nous Research switched Hermes Agent's default search to it and made it free.
At the site of the Bundibugyo Ebola outbreak in the Democratic Republic of the Congo, the WHO Regional Office for Africa, CEPI and the DRC's National Institute of Biomedical Research are using Claude. A situation report that took a full day now takes under an hour, and in a fight against a disease with no vaccine, what does that time change?
At Connect 2026, Meta unveiled a VR device shaped like a pair of glasses and Muse Charm, a device built for its agent. It also revealed a plan to bring its personal AI agent Muse to AI glasses, add more shopping and work partners, and make money from transaction fees.
Project Swap sent Claude-powered agents for 201 employees to haggle over books on their behalf. Reading people's tastes mattered more than trading skill, and the choice of model moved outcomes more than the instructions did.
Sam Altman called at the UN Security Council for national and international frontier AI standards and an incident reporting system. At the same meeting the US government rejected global AI governance.
Google Research has unveiled an AI video co-director that plans and fixes minutes-long videos on its own on top of Gemini and Veo. Four frameworks target characters, backgrounds and props that change from cut to cut.
Google has disclosed a server-side memory design for Private AI Compute that encrypts per-user memories in the cloud. The 15-page technical brief spells out how persistence raises the stakes — and what Google is doing about it.
ChatGPT Ads opened in seven new markets from Indonesia to Taiwan, taking the product past 60 countries. This time advertisers can run campaigns themselves in Ads Manager from the first day.
Indian voice and chat agent company Ringg handles more than 7 million calls a month with OpenAI models. Routing each task to a different model cut its model costs by about 90%.
Legal AI company Harvey uses GPT-6 Astra to bring more context into drafting and produce better-structured legal documents. The case study also introduces a memory panel that brings lawyers' preferences into the draft.
About 950 agents spent 21 hours combing a DNA database and turned up a repeat structure that looks like CRISPR. Humans handed over just one opening prompt — and still don't know what the system actually does.
Meta showed off a real-time video model at Connect 2026 that gives its personal agent Muse facial expressions and gestures. Voice and video come back together in roughly 870 milliseconds after a person finishes speaking.
Anthropic now lets threads in Claude Code projects run on your own computer, six days after launching Projects as a cloud-only beta on September 17.
After adding GPT-6 Astra, video-editing startup invideo saw its color-correction success rate rise about threefold. A few editors built about 50 custom effects in a single day.
ChatGPT Voice can now use email, calendar, and Slack plugins and runs on GPT-6 Astra, Sol, and Luna. In ChatGPT Work on web and mobile, you can also build documents and sites by voice.
OpenAI is working with Ukraine's Ministry of Digital Transformation to open its Daybreak cyber defense program. The focus is defending civilian infrastructure such as hospitals, energy and telecommunications.
Google has released Gemini 3.8 Flash TTS, which designs voices from written descriptions, and Flash-Lite TTS for high-volume work. Voice replication from a 30-second sample comes with verbal consent checks and SynthID watermarking.
Web research infrastructure startup Parallel says GPT-6 Astra cut the time and cost of its long research tasks in half. More targeted searching, not bigger reasoning, was named as the reason for the speedup.
A single document sets out four priority areas for deeper outside assessment and seven principles the work should follow. Claims are pre-registered before assessment, and assessors can note in their reports when the company has requested redactions.
OpenAI fixed the cache-discount reuse window at 30 minutes and shipped a hit-rate dashboard alongside a miss-diagnostic tool. Because cache writes cost 1.25 times the input price, hit rate is now directly tied to cost.
Anthropic has released Opus 5.5, the first model in its new Claude 5.5 line. It matches Claude Fable 5.1's performance on most tasks while running 40% cheaper than Opus 5, and cybersecurity-related requests are automatically routed to Opus 4.8 instead.
