大爱 AI 日报 AI Daily 2026/11/13: Anthropic Plans Nasdaq IPO Under Ticker ANT

Today’s Summary

Anthropic's IPO filing is public, planning to issue approximately 825 million shares for Nasdaq trading on November 19.
From Gemini's speed gains to OpenBot's wide release, models and open-source agents are pushing in the same direction, all competing for everyday developer workflows.
Today: try Gemini 4 Hydrogen's coding experience first, then wait for ERNIE 5.5's open-source release to verify its reasoning.

🔥 Today’s Top 10 Focus

1. Anthropic Plans Nasdaq IPO Under Ticker ANT

IPO filing is now public. The Anthropic official IPO page shows the company plans to issue approximately 825 million Class A shares at a price range of $70–$78 per share. Pricing is expected on November 18, with trading beginning November 19, subject to regulatory approval and market conditions. Q3 2026 revenue was $17 billion, with an adjusted operating margin of positive 15% and a net loss of $23.3 billion, of which $21.9 billion came from non-cash valuation changes. The company will retain its public benefit corporation status, with the founders’ entity controlling approximately 50.1% of general voting rights through Class F shares.

2. Google Launches Gemini 4 Hydrogen Focused on Everyday Speed

A lightweight addition to the Gemini 4 family. Google AI’s official tweet announces that Google has launched Gemini 4 Hydrogen focused on everyday speed , running 2.5× faster than Gemini 3.8 Flash. It surpasses Argon on the Artificial Analysis Intelligence Index at the default thinking level, while also supporting adjustable thinking levels. It is designed to work in tandem with Carbon inside Antigravity 🔧, and everyday developers should pay attention to real-world deployment performance.

3. Fei-Fei Li Releases Atlas Unified World Model

World models move toward robot training. According to 36Kr, Fei-Fei Li gave a talk at CoRL 2026 and released the Atlas unified world model , which uses a multimodal autoregressive diffusion Transformer architecture. This was Fei-Fei Li and Lisa Su’s first public appearance together since AMD announced its acquisition of World Labs for approximately $8.2 billion. Fei-Fei Li disclosed that Atlas is being tested for its practical value in robot simulation training, capable of generating and reconstructing three-dimensional environments.

4. ERNIE 5.5 Flash Preview Weights to Be Open-Sourced Before November 25

Baidu ERNIE’s new version makes its debut. According to AIBase, ERNIE 5.5 Flash Preview weights will be open-sourced before November 25 , with 280B total parameters and 14B active parameters, supporting a 1M ultra-long context. The weights are planned for official open-source release before November 25; the current version is a mid-stage post-training release. Demos for game development and 3D modeling scenarios look promising, but inference efficiency and stability will need to be verified once the official version is live.

5. GPT-6.1 SOL Benchmarked at 223 tok/s Output Speed

A noticeable jump in generation speed. A V2EX user’s post with screenshots benchmarking GPT-6.1 SOL at 223 tok/s shows an output speed of 223 tok/s. This is a personal benchmark, not an official figure, and results will vary by task and network conditions. For developers who need real-time interaction or batch generation, this number is worth keeping in mind.

6. Software Engineers’ Skill Sets Are Rapidly “De-technicalizing”

Technical moats are eroding. In an article shared and translated by Baoyu, the author describes concrete shifts in job roles : writing user stories for a model now outperforms writing rigorous technical specifications. The skills required for the job are shifting from technical to “human-oriented”, with a steady decline in demand for deep technical knowledge. The author advises actively seeking out areas where genuinely hard problems still exist, rather than clinging to old skills that AI is already covering.

7. The AI “Curse of Knowledge” Can Be Worked Around with Multi-Agent Collaboration

The smarter a single agent gets, the harder it is to simulate an ignorant perspective. In an article posted on Jike, Pinglun Shi compiled practical thoughts on using multi-agent architectures recently . The core problem is that once a single agent accumulates a large body of knowledge, it struggles to think from a beginner’s perspective. Multi-agent systems address this through role division, distributing different cognitive starting points across agents. These insights, drawn from real-world use, are a useful reference for developers designing agent workflows.

8. Chrome Team Releases Modern Web Guidance Toolkit

A web standards library installable from the command line. Gorden Sun’s tweet introduces the Chrome team’s modern-web-guidance web standards toolkit : after a single terminal command to install it, AI can automatically retrieve and apply standard examples. Testers report that responsive layouts become cleaner once it’s added, with accessibility and passkey login supported out of the box. The repository is at GoogleChrome/modern-web-guidance on GitHub.

9. Youdao Open-Source Office Agent Updated with Team Token Management

Team token management feature has been updated. According to an introductory article published on WeChat by Jiamu Weilai, the Youdao team’s open-source office agent has added a team-edition token management feature, along with several application case studies. The source is a social media roundup post; for authoritative repository details, refer to the official open-source page. This is worth noting for developers who need to manage and control AI call costs within a team.

10. Over 100 Data Centers Now Built or Under Construction in Ulanqab

AI computing power concentrates in inland green-energy low-cost zones. Ruan Yifeng, in Weekly Issue 419 , cites a blogger’s on-site visit: Ulanqab now has more than 100 data centers, with virtually every major IT and computing company establishing a presence there. An average annual temperature of 4.3°C provides natural cooling advantages, and abundant green energy makes this prefecture-level city a strategic hub for computing power. Practitioners who follow AI infrastructure siting and the green computing trend will find this firsthand account a useful reference.

⌘ Top Open-Source Projects

QwenLM/qwen-code: An AI Coding Agent for the Terminal

Another strong contender in terminal coding agents. QwenLM/qwen-code gained 151 stars today , bringing its total to 29,809, written in TypeScript. The project runs an AI coding agent directly in the terminal, handling code generation, modification, and debugging tasks. It is worth trying for programmers who prefer command-line workflows and want to embed AI into their local development environment.

CopilotKit/OpenBot: Every AI Colleague Has Its Own Computer

Decision before each action, log after each action 🔒. CopilotKit/OpenBot gained 298 stars today , bringing its total to 11,605, making it today’s highest-gaining AI project on the daily chart. Each agent is equipped with its own dedicated browser, files, and tools, and supports any AG-UI agent protocol. Suitable for teams that need to build auditable multi-agent workflows.

ashhart/TensorFold: A Cross-Platform LLM Inference Engine

Inference runs on Metal, CUDA, and Vulkan. ashhart/TensorFold gained 54 stars today , bringing its total to 1,938, written in Zig. It covers Apple Silicon, NVIDIA GPUs, and cross-platform scenarios, with the goal of unifying LLM inference backends across different hardware. Worth trying for developers who need to self-host LLMs across multiple hardware platforms.

❓ Related Questions

With Gemini 4 Hydrogen’s speed improvements, what should everyday developers try first?

Start by quickly running a prototype conversation in Antigravity. According to the introduction of Google’s lightweight new model , Hydrogen is 2.5× faster than Flash, making it well-suited for high-frequency iteration. To try it in Antigravity, check out 爱窝啦·AI账号店’s Gemini Pro ready-to-use account including Antigravity — this is a new ready-to-use account, not a renewal of an existing one.