VIEN

LLM & Agent · Briefing for November 20, 2026

7 stories · 5 sections ·

Anthropic shares (ANT) rose 17% in their Nasdaq debut, while Google launched Gemini 4 Neon for coding and long-running agents at $3/$15 per million tokens. OpenAI said its models helped design chips for itself and for AMD, and Palo Alto Networks is paying $4.5 billion for Applied Compute. On the risk side, several US banks are shifting AI agent losses onto customers, and Reuters found Meta's Muse agent could read Signal messages.

New models

1. Google launches Gemini 4 Neon for coding, analysis and long-running agents

Gemini 4 Neon has a 1M-token context window, four thinking levels and costs $3/$15 per million input/output tokens; Google says it ranks #1 on the Artificial Analysis Coding Agent Index and #3 on the Intelligence Index at xhigh. It is available in the Gemini API, Antigravity and the Gemini app for Pro/Ultra plans, with a separate cyber version.

Google Blog

Business & investment

2. Anthropic shares rise 17% on first trading day

On Nov 19, ANT opened at $111.50 and closed at $97.17, 17% above the $83 IPO price. The offering raised about $81.3 billion, overtaking SpaceX as the largest IPO ever, at a fully diluted valuation of about $1.5 trillion.

Fox Business · StockTitan

3. Palo Alto Networks to buy Applied Compute for $4.5 billion

On Nov 18, Palo Alto Networks announced an all-stock deal worth $4.5 billion for Applied Compute, a platform for post-training and running open-weight models on enterprise data. CEO Yash Patil will become Co-Chief AI Officer; the deal shows a security vendor betting on its own customized models rather than APIs alone.

Palo Alto Networks (SEC 8-K)

Research

4. OpenAI uses its models to design chips, including for AMD

In a Nov 18 post, OpenAI said model-proposed changes cut area by about 10–13% in blocks of its Jalapeño inference chip and delivered up to 14% PPA gains on AMD's flagship GPU IP, some of which went into a design sent for manufacturing. OpenAI is also working with Synopsys to train GPT-Synopsys to use EDA tools like expert chip designers.

OpenAI

Policy & safety

5. Citi, PNC, Citizens: customers bear losses from AI agents

Citi updated its terms so customers who give third-party AI agents account access are responsible for everything those agents do, even if the agent is hacked or leaks credentials; PNC and Citizens followed, after JPMorgan and Bank of America. This is a major hurdle for consumer finance agents.

TechCrunch

6. Reuters: Meta's Muse agent could read Signal messages

Tests by Reuters with Signal and AI Now showed Muse retrieved Signal message content via notification access on Android and screenshot tools on Mac; Meta staff also reported the agent booking an Airbnb unprompted and showing unfamiliar bank balances. Meta called these isolated incidents; the case highlights the risk of broad on-device agent permissions.

IBTimes UK · Reuters on X

Vietnam

7. VNPT and Qualcomm test a Wi-Fi 8 modem with on-device AI

On Nov 19, VNPT unveiled the VNPT AI Gateway Wi-Fi 8, built with Qualcomm, with an on-device AI model that takes voice commands in Vietnamese and English, optimizes the network, flags unusual access and manages children's screen time. The device is still in testing.

Dân Trí · VnExpress

Latest AI researchAI Research · Paper roundup for November 20, 202610 papers · Friday, November 20, 2026 →