The AI Spotlight Follow
Nvidia's Board Met About Its CEO, Anthropic Priced a $1.4 Trillion IPO and Google Shipped Carbon: Inside AI's Wildest Week
Model Wars

Nvidia's Board Met About Its CEO, Anthropic Priced a $1.4 Trillion IPO and Google Shipped Carbon: Inside AI's Wildest Week

In five days, the AI industry saw a leadership crisis at its most valuable chipmaker, one of the biggest IPOs ever set its price range, and three frontier labs ship or tease new models. Here is every story that matters, what is confirmed, and what is still rumor.

Advertisement

Contents

Some weeks in AI bring one big launch. This one brought a corporate thriller, a trillion-dollar stock offering and a pile of new models, all before Thursday.

The world's most important chipmaker spent the week weighing whether to remove the founder who built it. The company behind Claude launched an IPO roadshow that could value it at up to roughly $1.4 trillion. Google pushed its strongest model to the public and introduced a new kind of AI that makes decisions without writing a single word.

And that was only the headline layer.

Here is the full picture, sorted into what is confirmed and what is still rumor.

Advertisement

Google's element-named model machine

Google's Gemini 4 generation is named after chemical elements. Argon launched last month as the workhorse. Carbon arrived this week as the heavy thinker. Testers on public model-comparison arenas have also reported anonymous checkpoints under names like Hydrogen and Krypton.

Treat those last two carefully. Google has not announced a model called Krypton or Hydrogen. What testers describe is an early checkpoint, sometimes labeled "Krypton A," that appears to think for roughly twice as long as Carbon and produce roughly twice as much code per prompt. One tester's zero-shot 3D scene of a historical sailboat reportedly took more than 22 minutes to generate. Another prompt, a detailed SVG of a pelican riding a bicycle, showed more moving parts and finer detail than Carbon's attempt, though observers still rated Anthropic's models ahead on web design.

Google's Logan Kilpatrick fed the speculation with a short line that has been widely shared: Carbon "is a great model, but the team is still cooking."

Testers also report a possible video model nicknamed "Bouncy Boore," with early outputs said to support 4K, roughly 10-second clips that can be extended up to three times, and multi-image input. None of that is official, so treat it as a sighting rather than a spec sheet.

Gemini 4 Carbon, put to work

Carbon itself is very real. Google made it available on November 10, 2026, calling it "our strongest model with frontier performance in cybersecurity capabilities and long-horizon agentic workflows."

The pitch is endurance. Rather than answering fast, Carbon is designed to "explore thoroughly, reason deeply, and run extended verification loops," checking and refining its own work across many steps. That matters for long coding jobs and for hunting security vulnerabilities.

Key details, straight from Google's documentation:

  • Price: an introductory $2 per million input tokens and $10 per million output tokens through November 30. From December 1, it rises to $4 and $20, with 95% off cached input.
  • Context: a 1 million token window, with thinking levels of low, medium and high. The default is medium.
  • Where: the paid Gemini API in AI Studio, Gemini for business, and, for Google AI Ultra subscribers, the Gemini app, AI Mode in Search and Antigravity.
  • Scientists: Google is offering eligible researchers six months of free access in Antigravity's top usage tier.

That default-to-medium choice lines up with what hands-on testers report: high reasoning is only marginally better on many tasks, and medium saves a lot of tokens.

What the outputs looked like

The demos circulating this week give a feel for the model's personality:

  • A 3D scene built entirely in custom C++ with no reused assets, which took about two rounds of prompting and drew praise for its lighting and atmosphere.
  • A head-to-head against OpenAI's GPT-6.1 Sol on the same prompt at maximum reasoning. Per the tester, GPT-6.1 Sol ran about 30 minutes, used around 5.3 million tokens and cost about $1.40. Carbon ran about 40 minutes, output about 6.2 million tokens and cost about $2.90. Carbon's result looked more detailed; its interface was clunkier.
  • A community-built Call of Duty-style clone, and a product-launch animation for a fictional lighting instrument that exploded into components and cycled through finishes.

The consensus take: Carbon is a strong, competitively priced model for complex coding, still a step behind Anthropic's best on front-end polish. Many users also complain that consumer Gemini plan limits run out quickly.

The Decisions API: AI that answers without writing

The most quietly important Google launch this week may be the Gemini Decisions API, powered by a new model called Gemini Decision 1 (gemini-decision-1).

Normal language models answer by generating text, token by token. Decision 1 does not generate text at all. You hand it one input, which can be text, images, audio, video or a PDF up to about 256,000 tokens, plus as many as 1,000 questions. It scores every question in a single forward pass and returns calibrated probabilities with a confidence score.

Three question types are supported:

  • Yes/no checks, returning a probability between 0.01 and 0.99
  • Multiple choice across 2 to 255 options
  • Rubric scores across ordered levels

Google says answers arrive in 150 to 250 milliseconds end to end, at $0.10 per million input tokens with zero output tokens billed. Google's demo apps include "Doorstep Parcels," which judges delivery photos for questions like whether a package is damaged or left in a sensible spot.

Why it matters: agents make thousands of small judgments. Should this email go to billing or support? Is this photo acceptable? Which tool should run next? Using a big reasoning model for each of those is slow and expensive. A dedicated, near-instant judge could make agent workflows far cheaper. Google is already using it inside a Stitch design experiment that generates multi-screen app mockups in seconds.

The Nvidia leadership crisis

The week's most shocking story had nothing to do with model quality.

On November 8, Nvidia's board announced an independent investigation "in light of recent reporting that founder and CEO Jensen Huang has been the target of a foreign intelligence effort." The board said it had "no reason to believe that either national security or NVIDIA secrets were compromised," and that Huang supported the review.

Events then moved fast:

  • Reuters reported that Huang acknowledged in a memo to employees that he had been unfaithful and took "full responsibility." The woman involved is alleged to be a Chinese spy. China has denied the allegations.
  • The Washington Post and others reported a second affair, this time with an Nvidia executive who reported to him.
  • The Wall Street Journal reported the board was preparing to decide whether to remove him. Semafor reported some directors favor CFO Colette Kress as interim CEO, and Bloomberg reported an outside search firm was looking at successors.
  • Reuters reported that Nvidia's stock slid about 7% in early Frankfurt trading after the Journal's report.

As of this writing, no removal has been announced. Under Nvidia's bylaws, directors can remove the CEO "with or without cause" by majority vote. Huang co-founded Nvidia in 1993 and has led it ever since, so any change at the top would be one of the biggest leadership shocks the tech industry has seen. Some reports also describe strong employee support for keeping him; that has not been formally verified.

For AI, the stakes are practical. Nvidia supplies most of the chips that train and run frontier models. Leadership uncertainty touches product roadmaps, export-control politics and the market's appetite for AI spending.

Anthropic's IPO and its office-suite push

Anthropic's filing this week makes its IPO concrete. The company launched its roadshow on November 10, offering 825 million Class A shares at $70 to $78 and applying to list on Nasdaq as ANT.

The numbers, from the amended S-1 and reporting by CNBC and others:

  • At $78, the base offering is worth about $64.35 billion before fees.
  • CNBC calculated a fully diluted valuation of up to about $1.405 trillion, up from $965 billion in a May funding round. Other outlets using different math put it nearer $1.1 to $1.33 trillion.
  • AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC agreed to buy $26 billion of non-voting stock in a private placement at the IPO price.
  • An underwriter option for up to 123.75 million more shares could take total proceeds toward roughly $100 billion at the top of the range, according to The Guardian.
  • CNBC reported plans to price on November 18 and trade on November 19, with roughly 20% of shares aimed at retail investors.

Reports also say the order book is heavily covered, and some say bankers have discussed pricing higher. Those are reports from people familiar with the deal, not filings.

New tools aimed at Microsoft and Google

Anthropic also launched three collaborative work products:

  • Claude Whiteboards (beta): describe a messy idea and Claude organizes it into an editable visual board your team can work on together.
  • Claude Sheets (beta): generate working spreadsheets, like budgets, trackers and project plans, by describing them, then keep editing as a team.
  • Claude Data (generally available, formerly Dashboards): connect company data to build interactive dashboards, data stories and even narrated videos you review before sharing.

It also introduced interactive connectors that open third-party apps such as Adobe, Atlassian and Box inside Claude, so you can work in them directly or let Claude help.

And it gave every Pro, Max, Ultra and Team subscriber one free full usage reset across all models and limits, usable until December 10 at 2 p.m. Pacific, after fixing issues that had consumed more usage than intended.

The Fable-versus-Opus subscription math

One of the most useful posts of the week came from a well-known developer-creator who logged over 1.1 million real Claude Code requests on the $200 Ultra plan. His headline: the plan delivered roughly $15,000 of API-equivalent Opus 5.5 usage a month, but only around $3,000 of Fable, a roughly five-fold gap.

Two reasons stood out in his analysis. Fable can use only 50% of the weekly limit, so a Fable-only user effectively loses half the plan. And Fable burns through limits faster per dollar than Opus even before that cap.

Fable 5.2 is cheaper than 5.1 at the API level, he noted: about $8 per million input tokens (down from $10) and $40 per million output (down from $50).

His most interesting metric was work shipped per week. Opus alone gave him about 1,200 turns and 230 threads. Maxing out Fable first and then switching to Opus dropped him to about 900 turns but slightly more finished threads, because Fable more often finishes a job in one prompt. He cautioned that these are early numbers skewed by benchmarking. The practical rule: Fable for hard, one-shot problems; Opus for the long grind; Sonnet and Haiku for simple tasks.

OpenAI: GPT-6.2 whispers and a business directory

OpenAI watchers spotted a change in API documentation referring to "models 6.2 and later," plural. That has fueled speculation that a whole GPT-6.2 family, possibly including variants like Astra and Sol, is coming together. A prominent OpenAI employee's teaser about a "big day" added fuel. Nothing is announced. Treat release-date chatter as rumor.

What did launch is Businesses in ChatGPT, in beta. Companies can claim a profile, update details and upload product feeds with current prices and availability. Eligible US e-commerce and online-service businesses can also publish a Business Agent that answers customer questions inside ChatGPT, clearly labeled and separate from ChatGPT's own answers. It cannot complete purchases or bookings. Aggregated analytics show mentions and clicks without exposing private chats.

OpenAI also folded a finance feature into the ChatGPT desktop app for US users, and updated its personal agent so one agent can use several of your computers, for example building and testing an app on a desktop while you keep coding on a laptop.

China's challengers: Qwen4-Max and MiniMax H3-Image

Alibaba opened a preview of Qwen4-Max on November 11. It runs until November 25 (Beijing time) and is available through the Qwen app, Qwen Studio, QwenWork, Qoder and a paid Token Plan subscription, not pay-as-you-go API billing. Alibaba's documentation lists a 1 million token context, up to 393,216 output tokens, image and video understanding, tool calling and compatibility with OpenAI and Anthropic APIs. Alibaba has not published benchmark scores.

In one tester's comparison, Qwen4-Max finished a koi-pond scene in about 30 minutes, versus about 70 minutes and $6.30 for MiniMax M3.1. A 3D Sagrada Família attempt was cut off by a token limit but impressed for a partial run. Separately, Alibaba made its Qwen Code 1.0 preview source-available, though outside code contributions will not be accepted until after December 2026.

MiniMax H3-Image, released November 10, is an open-weight model that generates native images around 16 megapixels, up to 4096 by 4096, and supports editing with up to 16 reference images. One important caveat: its community license excludes the EU, UK, US and South Korea, and larger commercial use needs written permission.

Agents everywhere

This week made one thing obvious: every company now wants its AI to act as a coworker.

  • Google Antigravity /swarm: one command launches a coordinated team of agents for planning, coding and verification, communicating through a shared message board. You can mix models across roles.
  • Grok Bot 1.0: SpaceXSI's update adds shareable AI teammates that work inside Slack and Microsoft Teams, learn team context and follow up on unfinished work. Bots can also join Google Meet, Zoom and Teams calls and produce summaries and action items. Team bots require Team or Enterprise plans.
  • Collaborative coding workspaces: coding tools are adding multiplayer projects where teammates and their agents share progress in one place.
  • Google Labs Tabletop: an experiment that turns a group idea into a shared interactive tool, such as a road-trip map, a dinner decision wheel, a tournament bracket or a seating chart, with no code. It has a waitlist for US users 18 and over.

Grok 4.8 and a rebuilt stack

Elon Musk said this week that Grok 4.8 has finished pre-training and is now in reinforcement learning, with evaluation scores "starting to climb." The delay, he said, came from dumping JAX and SGLang and rewriting the entire RL and inference stack from scratch in C and C++, plus heavy cleanup of fine-tuning data.

He said a soft rollout to Grok Bot users was planned for later in the week, and that 4.8 "will bear no resemblance to 4.7." He also admitted it is "not yet incredible," but argued the foundations are now right. Those are a CEO's claims; judge the model when it ships.

Robots on the factory floor

Google DeepMind announced Gemini Robotics 2.5, developed with help from Gemini 4 Argon. In a pilot with Apptronik's Apollo 2 humanoids at Mercedes-Benz plants, the robots picked metal brackets from bins and placed them in sequencing trays, a job called kitting. DeepMind reports a 96% success rate. Critics note it did not publish trial counts or cycle-time comparisons, so treat the figure as promising, not proven.

The "age of miracles" claim

To close the week, Anthropic CEO Dario Amodei repeated one of his boldest predictions in a widely shared clip: that cancer, Alzheimer's, heart disease and mental illness could become "relics of the past" in less than a decade, ushering in "an age of miracles" if the risks are handled carefully.

It is a prediction, not a finding. Biology moves at the speed of clinical trials, regulators and manufacturing. But it explains why investors are lining up for his company's shares.

Why this matters

Three forces collided this week:

  1. Capital at unprecedented scale. Anthropic's raise, plus reporting that it plans enormous compute spending, shows how expensive the frontier has become.
  2. Concentration risk. One executive's personal crisis moved hundreds of billions in market value, because so much of AI runs on one company's chips.
  3. Commoditized intelligence. Carbon's pricing, Decision 1's $0.10 rate and open-weight releases from China keep pushing the cost of capability down.

For users, that means more powerful tools at lower prices. For the industry, it means more fragility in fewer places.

Practical takeaways

  • Try Carbon before December 1. The introductory API price doubles after November 30.
  • Use your Claude reset wisely. Save it for a heavy week before December 10.
  • Route by task. Fable for hard one-shot problems, Opus for long sessions, smaller models for routine work.
  • Prototype decisions cheaply. If your agent spends tokens on yes/no judgments, test the Decisions API's free tier.
  • Claim your business profile. If you sell online in the US, check Businesses in ChatGPT so your product info is accurate where customers ask.
  • Read licenses. "Open-weight" does not always mean usable everywhere. MiniMax H3-Image excludes several major markets.

Counterpoints and risks

  • Rumor inflation. Krypton, Hydrogen and GPT-6.2 dates are unconfirmed. Leaks often arrive weeks or months before, or instead of, launches.
  • Self-reported numbers. Benchmark claims from labs and single-user cost tests are useful signals, not independent audits.
  • IPO hype. Oversubscription reports come from anonymous sources. Valuations depend on how you count options and unvested shares.
  • The Nvidia story is unresolved. Investigations take time, and early reporting can change.

FAQ

Is Gemini Krypton a real Google model?

Not officially. Testers report anonymous checkpoints using the name on public arenas, and a Google lead hinted the team is "still cooking," but Google has announced only Argon and Carbon in the Gemini 4 family so far.

How much does Gemini 4 Carbon cost?

$2 per million input tokens and $10 per million output tokens through November 30, 2026, then $4 and $20 from December 1, with 95% off cached input.

When does Anthropic's IPO happen?

CNBC reported plans to price on November 18 and begin trading on November 19 under the ticker ANT. Those dates come from people familiar with the deal and could change.

Has Nvidia removed Jensen Huang?

As of this writing, no removal has been announced. The board opened an investigation on November 8 and, according to multiple reports, met to weigh his future.

What is the Gemini Decisions API used for?

Fast, structured judgments: routing requests, checking policies, scoring candidates against rubrics and picking tools, returning probabilities instead of text in 150 to 250 milliseconds.

The bottom line

This was the week AI stopped looking like a tech story and started looking like the whole economy in miniature: trillion-dollar listings, boardroom drama, geopolitics and a flood of cheaper, smarter tools.

The models keep improving on schedule. What changed is the scale of everything around them. Keep an eye on three dates: November 19 for Anthropic's debut, November 30 for Carbon's price change, and whatever Nvidia's board announces next.

Sources

AI Data Centers Are Failing the Kitchen-Table Test. The Real Fight Isn't About Water, It's Your Power BillUp next · 15 min readAI Data Centers Are Failing the Kitchen-Table Test. The Real Fight Isn't About Water, It's Your Power BillData centers went from invisible warehouses to the most fought-over buildings in America in about three years. The backlash is not really about AI, or even water. It is about a building that uses a city's worth of electricity for people who live somewhere else, and who ends up paying for it.Keep reading →