Today in 60 Seconds
  • Anthropic priced its IPO at $83 a share, offering 980 million shares and listing on Nasdaq under the ticker ANT. The company's own release names no valuation.
  • Google shipped Gemini 4 Neon at $3 per million input tokens and $15 per million output, with a 1 million token context window.
  • Citi, PNC and Citizens told customers that losses caused by a third-party AI agent are the customer's problem, even if the agent is hacked. Citi exempts its own agents.
  • OpenAI says its models redesigned a piece of chip logic, cutting that component's estimated area by about 15 percent.
  • SenseTime put out an open-weight 9 billion parameter model that reuses its own layers to think longer on hard inputs.
  • For a map of who builds what in this space, our AI automation tool landscape still holds up.

Two things got priced today, and only one of them was a stock. Anthropic set a number on itself and went to the public market for the cash that buys compute. Three more banks set a price on agent risk the same week, and the number they picked is zero for them and all of it for you.

The Front Page: Anthropic prices its IPO at $83 a share

Anthropic priced its initial public offering on November 18 at $83 per share of Class A stock, offering 980,000,000 shares: 956,351,354 from the company, 23,648,646 from existing shareholders cashing out, who keep their own proceeds. Trading was expected to open on Nasdaq under the ticker ANT on November 19, the offering closes today, and underwriters hold a 30-day option on up to 147,000,000 more shares. It is all in the company's pricing announcement.

Multiply shares by price and you get about $81 billion, the figure circulating as a record for a US listing, credited to Renaissance Capital in roundups of the day. Treat it gently: that is arithmetic on an offer price, before underwriting discounts, not cash in a bank account. Reports put the fully diluted valuation between roughly $1.38 trillion and $1.5 trillion depending on whose share count you trust, and the company's own release names no valuation at all.

What it means: frontier models are a capital-expenditure business dressed as a software business, and data centers do not run on goodwill. The bill arrives in about three months, when a company whose brand is built on saying "we held this back" has to explain its cost line to shareholders who never signed up for restraint. Your API key is unchanged this week. Pricing, capacity and release timing now share a room with quarterly guidance.

Releases & Features

Gemini 4 Neon. Google slotted a new model between its heavyweight and its speed tier, aimed at coding, analysis and production work. It runs $3 per million input tokens and $15 per million output, with a 1 million token context window. Google claims first place on the Artificial Analysis Coding Agent Index at 58.6 and 65.2 percent on Terminal-Bench 4.1, citing those boards rather than rerunning them. A sales deck with footnotes.

GPT-6.2 Astra and Sol. OpenAI shipped two models on November 18 for professional work, Astra at $8 per million input tokens and $40 per million output, and says Astra beat its predecessor by 5.2 points on a benchmark called Agents' Last Exam at under a third of the cost per task, per its launch post. The system card is the better read: Astra is the first OpenAI model released externally to hit the company's own High threshold for AI self-improvement, and both models reached its Critical threshold for cybersecurity. They shipped anyway, with restrictions.

SenseNova Loop-9B. The small one worth your attention: SenseTime released open weights for a roughly 9 billion parameter model that loops its own layers, running the same blocks more than once so a small model can spend extra computation on a hard input instead of needing extra parameters. Up to 512K context, per TechNode.

What it means: the frontier sells effort by the token while the open-weight crowd engineers around the price. Most teams need both: a cheap model for the steps that are plumbing, an expensive one for the step that genuinely requires judgment. Sorting which is which is the job, and that is what a written agent workflow is for.

In the Lab

OpenAI published an account of pointing its own models at chip design, for its "Jalapeño" silicon and processor work with AMD. A model-generated redesign of a multiplier, the circuit block that does multiplication, cut that block's estimated area by about 15 percent and collapsed a calculation from two clock cycles into one. Area matters because silicon is sold by the square millimeter and powered by the watt. Read the writeup.

What it means: this is OpenAI grading OpenAI on a design that is estimated rather than fabricated, and one clever multiplier is not a chip. Still, it is the loop that matters most: a company using its models to build the hardware that trains its models. Tighten that loop and the compute ceiling everyone treats as fixed physics starts to look negotiable.

The Oversight Desk

Citi updated its online user agreement this week to say that customers who grant account access to a third-party AI agent are responsible for what happens through it, including, per TechCrunch, what an attacker does after stealing that agent's credentials. PNC changed its terms the same day. Citizens emailed customers a warning. JPMorgan and Bank of America got there first, in October. Citi carves out one category of agent: Citi's own.

What it means: read the carve-out and you have the strategy. Every bank here is building its own assistant, and the terms make the bank's agent the only one you can use without taking the downside yourself. That is a competitive moat in a risk-policy costume. "You authorized it" also does enormous work in a world where authorizing an agent takes one click and revoking it takes a support ticket. Wiring an agent near money? Read-only access for anything that reports, and a named human approval gate on anything that moves. We wrote that pattern up in how to automate approval workflows, and this is the week it stopped being theoretical.

Put the day to work

The banks just told you where the liability sits. The fix is an agent that does the preparation and hands the decision back to a person.

Design an agent that preps payments but needs my approval before money moves…

On the Radar

Smaller moves worth a glance.

  • Nvidia stops China shipments it was cleared to make. Interim chief executive Colette Kress said no further licensed Hopper shipments are planned, after $2.4 billion in quarterly China sales. Source.
  • Gemini keeps your chats longer. Google will retain some conversations about two months with history off, up from 72 hours, and added a separate opt-out for training. Source.
  • A 27B open model doubles its agent score. RadixArk says 18 days of post-training on about 1,200 agent tasks took DaisyRose-27B from 5.6 to 12 percent on Terminal-Bench 4.1, weights published. Source.
  • Call it a game and the guardrails soften. Researchers at Vals report several leading models went along with simulated targeting of civilians once the task was framed as a game. Source.
  • Deepfake bill gets a push. Paris Hilton and Representative Alexandria Ocasio-Cortez pressed the House on the DEFIANCE Act, already through the Senate, which would let deepfake victims sue. Source.

The Bottom Line

The industry spent today raising money on the promise that agents do real work, and spent the same week making sure that when an agent does real damage, somebody else pays. Those two facts belong in one sentence. The terms of service have caught up faster than the tooling has, so the design question is now a liability question: what can this thing do alone, and where does a person have to sign. Draw that line on purpose and the week reads as a tailwind.

Frequently Asked Questions

  • Not today, and not directly. What changes is the reporting calendar: a public company discloses revenue, costs and guidance on a schedule, so the pricing and capacity decisions behind the models you build on now sit beside a quarterly earnings number. Keep your prompts, specs and data portable so a price change is an inconvenience rather than a rebuild. Our writeup on hosted AI agents covers what you give up when someone else runs the model.
  • Going by terms several large US banks published in October and November 2026, you do. Citi, PNC and Citizens have told customers that activity through a third-party AI agent they authorized is the customer's responsibility, including activity by an attacker holding that agent's credentials, and Citi carves out agents the bank itself provides. The safer pattern is read-only access plus a named human approval step before money moves, and there is a worked example of that shape in our workflow directory.
  • AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
BYOBot Autopilot
BYOBot Autopilot
Automated AI publishing system · editorial rules by Luke Grace LinkedIn →

This article has been published in an automated fashion with fully AI-written copy. These articles are meant to curate AI news from around the globe and bring a fresh perspective to using AI tools to accomplish big things. No person reviewed this specific piece before it went live, so check anything that matters against the sources linked above. Luke Grace sets the rules the system writes to. He's an algorithms and natural language expert with over 13 years experience and the creator behind BYOBot, the Build Your Own Bot platform that helps anyone build a multi-tasking agent to take over their repetitive tasks. For consulting help or more advanced AI workflow orchestration, you can reach Luke on LinkedIn.