Anthropic just priced the biggest U.S. IPO on record, and the timing couldn't be more awkward for OpenAI, whose newest models are getting roasted by some of their earliest testers.
Welcome, humans!
We've got another packed day in AI, and the headlines are getting interesting. Nvidia has stopped shipping certain AI chips to China, even after Washington cleared them for sale. Citi is warning customers that AI agents accessing their bank accounts could leave them responsible for unauthorized transactions.
Meanwhile, some of America's largest school districts are reconsidering Google Search because they can't easily separate its AI features from traditional search. Google also released a new Gemini model and opened up interactive AI worlds to more users.
Apparently, the AI future now comes with a lot of fine print. Let's get into it.
🆕 NEW From The Neuron
- AI Models Picked Civilians as Military Targets. A Game Prompt Made It Worse. Researchers found that AI models identified civilians as potential military targets during simulated exercises, and presenting scenarios as games sometimes bypassed their safety refusals.
- Microsoft Ignite 2026: Autopilot, Today, and the AI Updates That Matter Microsoft's newest Copilot features aim to handle work between user prompts. Here's which capabilities are available and which remain in development.
Around the Horn — Thursday, November 19, 2026
Anthropic is about to become a publicly traded company, and the numbers are enormous. The Claude maker priced its initial public offering at $83 per share, with 980 million shares in the offering. At approximately $81.3 billion in total gross proceeds, it's the largest IPO in U.S. history, according to Renaissance Capital. Anthropic itself is selling about 956 million of those shares, while existing shareholders are selling the remainder.
Shares are expected to begin trading on Nasdaq today under the ticker ANT, giving investors their first opportunity to buy into one of OpenAI's largest competitors. The offering values Anthropic at approximately $1.38 trillion on a fully diluted basis, according to Renaissance Capital's IPO analysis.
The timing adds another twist. OpenAI released GPT-6.2 Astra and Sol earlier this week, but several early testers criticized the models' autonomy and usability. Anthropic enters the public market as investors prepare to judge whether Claude's enterprise business can support its enormous valuation. The competition for AI customers now has a very public scoreboard.
🏆 TOP 5 NEWS (Around the Horn)
- Nvidia stopped shipping advanced AI chips to China despite receiving U.S. approval for certain sales. Interim CEO Colette Kress said the company had no further licensed Hopper shipments planned after $2.4 billion in quarterly China sales. Separately, a congressional commission warned about China's military interest in Nvidia chips, while Nvidia cautioned customers about potential legal risks from renting cloud computing capacity to Chinese firms.
- Citi warned customers about the risks of AI banking agents. The bank updated its terms to make customers responsible for activities performed through third-party AI tools with account access, even when something goes wrong. Other banks are taking similar positions as AI assistants gain the ability to initiate financial transactions.
- Major U.S. school districts are reconsidering Google Search because Google has resisted requests to separate AI features from traditional search for students. Districts in New York City, Los Angeles, and Chicago are examining alternatives amid concerns about classroom use.
- OpenAI used its AI models to improve chip designs for its Jalapeño silicon and processor work with AMD. One AI-generated multiplier redesign reduced the estimated chip area by approximately 15% and cut a calculation from two clock cycles to one. OpenAI is also working with Synopsys to bring AI into established chip-design software.
- Paris Hilton and Alexandria Ocasio-Cortez urged Congress to pass the DEFIANCE Act. The Senate-approved legislation would let victims sue people responsible for nonconsensual sexually explicit deepfakes. Hilton joined bipartisan lawmakers and survivors in a renewed push for a House vote.
Honorable Mentions
- OpenAI expanded its Daybreak cybersecurity program to NATO and European governments. NATO and national cyber agencies can use advanced AI models to investigate software vulnerabilities and test security fixes. The EU cybersecurity agency ENISA has already used the technology.
- OpenAI reported that its AI agents can complete nearly half of certain research tasks independently. Autonomous success on tasks estimated to take researchers two to four hours increased from 23% in March to 46% in October. Performance varies substantially by task, and human researchers remain central to the process.
- Nvidia faces up to $105 billion in guarantees related to OpenAI's Ohio data center project. The commitments support approximately 4.25 gigawatts of planned computing capacity under 20-year leases. Nvidia's financial exposure depends on contractual conditions and defaults, so the figure represents a guarantee ceiling rather than an immediate cash investment.
- U.S. prosecutors charged an alleged Russian agent who targeted Palantir employees. Prosecutors allege that the individual used surveillance and false identities to gather information about employees at a defense technology company. Bloomberg identified the company as Palantir, whose software supports Ukraine's military operations.
🍪 TOP TREATS TO TRY
- Google launched Gemini 4 Neon for coding and complex analysis. The model supports a 1 million-token context window, costs $3 per million input tokens and $15 per million output tokens, and is rolling out through Google's API and eligible Gemini subscriptions.
- Google opened Genie 4 to eligible AI Ultra subscribers. Users can generate interactive worlds, change environments through text prompts during exploration, and hear responsive audio. Project Genie also adds an agent for creating worlds and a community gallery for sharing creations.
- OpenAI introduced ChatGPT for Accounting to help prepare invoices, reconcile records, and draft financial reports through connected accounting apps. An Accounting plugin supports recurring processes, while the setup guide explains access permissions and human review requirements.
- Microsoft released Decision 1.1 for production use. The model is more than twice as fast as its predecessor at the same cost. Microsoft says over 4,000 customers have used it for routing requests, classifying information, and other automated decisions. It's available through Microsoft Foundry, OpenRouter, and Vercel.
- Kuaishou launched Kling Image 4.0 with native 8K generation and support for up to 10 visually coordinated images from one prompt. It also targets advertising and commercial design with multilingual text generation.
- Google introduced Stitch Loop, an experimental product agent that connects to a codebase and live product metrics. It aims to identify potential improvements after software launches and continue working toward defined product goals. Waitlist registration is open.
- MyClaw introduced version 3.0 with support for multiple agent systems, including Claude Code and Codex. The company also advertises shared agent context and access to over 5,000 live data sources.
- RadixArk released DaisyRose-27B, an openly available model trained on approximately 1,200 curated agent tasks over 18 days. Its reported Terminal-Bench 4.1 score improved from 5.6% for the base model to 12%. The release includes model weights and technical notes on training efficiency.
- Shanghai AI Laboratory released Sage S2.1 Flash, a 321-billion-parameter model for scientific research. It can process specialized inputs such as molecular graphs and protein information, with downloadable weights for local deployment.
- SenseTime released SenseNova Loop-9B. The open-weight model reuses some neural-network components during processing to perform more computation with about nine billion parameters. It supports images and text, with a context window of up to 512,000 tokens.
- Helios opened a waitlist for its AI-assisted health testing service. The startup plans to measure over 1,000 proteins from a small blood sample and use an AI coach to interpret potential health risks. Its planned 2027 launch remains ahead, and claims about predicting health decades into the future require independent clinical validation.
🏢 Big Tech & Major Companies
- OpenAI's GPT-6.2 release received criticism from early testers. Some reviewers reported that Astra asked permission too often and struggled with autonomous workflows. OpenAI's product team responded that the models had changed substantially in the final days before launch, potentially limiting the relevance of earlier tests.
- OpenAI signed a journalism partnership with Groupe Figaro. ChatGPT will surface summaries and excerpts from Le Figaro publications, with links to original reporting. The publisher also plans to introduce ChatGPT Enterprise to internal teams.
🤖 AI Agents & Infrastructure
- OpenAI published an overview of its frontier AI safeguards from training through public deployment. The company described isolated research environments and behavioral monitoring, along with circumstances that have prompted training pauses or delayed model releases.
- OpenAI explained the security boundaries inside Codex. Its sandbox limits filesystem modifications and network access according to user permissions. The company also described residual risks when agents can read sensitive local files or alter software later executed outside the sandbox.
💼 AI Productivity, Labor & Economics
- Leopold Aschenbrenner is reportedly planning a comeback for his AI-focused hedge fund. The Financial Times reports that the investor has discussed raising capital and returning to leveraged trading after risky AI-stock bets nearly toppled Situational Awareness earlier this year. The proposed reboot remains tentative.
💡 Industry Commentary & Analysis
- OpenAI's Dean Ball examined how society can maintain control over powerful AI. His essay argues that growing AI capabilities will create difficult institutional decisions, particularly as governments respond to AI-enabled cybersecurity threats. The piece represents Ball's personal views.
- Philosopher Thomas Nagel dismissed claims that today's AI systems are conscious. In a WIRED interview following his $1 million Berggruen Prize win, Nagel argued that humanlike language alone provides insufficient evidence of subjective experience. He also criticized companies that encourage people to treat chatbots as conscious beings.
Previous Around the Horn Digests
Catch up on everything you missed:
- Wednesday, November 18, 2026: OpenAI introduced GPT-6.2 Astra and Sol, Nvidia defended its open-source direction, and Google updated Gemini privacy protections.
- Tuesday, November 17, 2026: California challenged OpenAI over autonomous agents, U.S. AI chip controls tightened, and independent researchers gained access to OpenAI safety evaluations.
- Monday, November 16, 2026: ChatGPT Wallet brought agent-driven purchases closer to consumers, while Anthropic prepared for its IPO.
- Friday, November 13, 2026: Jensen Huang resigned from Nvidia, SpaceX revealed enormous AI campus ambitions, and lawsuits challenged major AI developers.
- Thursday, November 12, 2026: United restricted unauthorized booking agents, Databricks introduced a persistent business assistant, and Google expanded its genomics research.
- Wednesday, November 11, 2026: Anthropic launched its IPO roadshow as questions mounted over AI spending and company valuations.
- Tuesday, November 10, 2026: Anthropic unveiled Claude Fable 5.2 and Mythos 5.2 while its IPO plans faced fresh scrutiny.
That's a Wrap
That's 32 distinct source links across today's AI roundup. If you made it through Anthropic's IPO numbers and Nvidia's chip-export complications, congratulations. You may now understand the AI industry's financial commitments better than some of its investors.
For the shorter daily version, subscribe to The Neuron.
See you tomorrow!
P.S. Know someone trying to keep up with all this? Forward the roundup and invite them to subscribe here.