xAI put Grok Voice 3.0 on the Voice Agent API on October 14, a speech-to-speech model at the same eight cents a minute as Voice 2.0, with Artificial Analysis scoring it 86.9 on the speech-to-speech index. Google opened Deep Review to the public the same day: an agent that reads a technical PDF and writes a review, two complimentary passes a week, after a year of private conference pilots. Anthropic opened a Singapore office and signed an MOU with the Ministry of Digital Development and Information at the city's cyber week. Meta's Newsroom introduced Muse Wonder and Muse Wonder Full, agent models on the Meta Model API and Muse Code.
In the news
Grok Voice 3.0
xAI's October 14 news post puts Grok Voice 3.0 on the Voice Agent API, a first-party speech-to-speech model with sub-one-second latency at the same $0.08 per minute ($4.80 per hour) as Voice 2.0. Artificial Analysis scores it 86.9 on the Speech-to-Speech Index, ahead of Voice 2.0 at 81.3, GPT-Live-1 Astra Medium at 83.2, and Gemini 3.8 Live at 82.6. Agents already pointed at grok-voice-latest switched automatically. This is a voice model on the API, not the X Subscription consumer bundle and not Grok 4.7 on Amazon Bedrock. Grok 4.8 has not shipped.1
xAI says Tesla runs millions of in-car conversations a day and Starlink call centers take about 20,000 calls a day, with roughly 70 percent resolved with no human. Those volumes are the company's. Full Duplex Bench is 97 percent, τ-Voice is 71.6 percent (xAI says that is comparable with GPT-Live-1 Astra at about 40 percent lower latency), and EVA-Bench accuracy is 74 percent on ServiceNow's enterprise voice suite. Those Tesla and Starlink rates, the intelligence ranking, and the latency cut are vendor- or lab-reported.
Deep Review
Google Research opened Deep Review to the public on October 14 at research.google/deepreview. The agent reads a technical PDF or plain-text document and writes a review, two complimentary passes a week, with a Gemini API key unlocking volume and a Deep Review API. It now runs on Gemini 4 Argon. Uploaded documents are not used to train Google models and are stored one year for display in the product UI. The public launch follows private Paper Assistant Tool pilots on more than 16,000 opted-in manuscripts at STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027, decoupled from official conference review.2
Google describes a four-stage pipeline: document segmentation, adaptive budgeting, deep parallel verification, then global synthesis and grounding. Multi-file input is still coming. The 16,000-manuscript count and the 74,000-submission figure for ICLR, NeurIPS, and ICML this year are Google-reported.
Singapore office
Anthropic opened a Singapore office on October 14 and signed an MOU with Singapore's Ministry of Digital Development and Information at Singapore International Cyber Week. It is the company's fifth APAC office. Permanent Secretary Kai Fong Chng and Anthropic co-founder Ben Mann signed; Josephine Teo and Michael Sellitto witnessed. The framework covers AI safety, security, model evaluation, and research on adoption and impact.3
Anthropic names GovTech (Claude Code for more than 1,500 engineers), DBS, GIC, Grab's BriX sales assistant, Sea/Shopee, and Supabase among customers. Its Economic Index puts Singapore second of 121 countries in Claude.ai usage relative to population, about six times what population would predict. Grab's 40,000 hours a quarter and that six-times usage figure are Anthropic or customer-reported. In-region Bedrock inference is available in Seoul and Singapore, and the office is open and hiring.
Muse Wonder
Meta's Newsroom post dated October 12 and October 13 introduces Muse Wonder and Muse Wonder Full, built on Muse Spark. Muse Wonder is available through the Meta Model API and Muse Code. Wonder Full is for Muse Code subscribers. The company describes Active Memory, which carries relevant details across conversations, and Active Focus, which stays on a hard task, tries other approaches, recovers from errors, and checks in before important actions.4
Meta calls these its most capable models to date and lists expert domains in law, finance, and health, plus coding. A full eval report is promised, not posted. The research write-up is not up, and there is no public scoreboard in the Newsroom post.
