ThursdAIThursday, Nov 12, 2026Live from AI Engineer Code, San Francisco2:09:04Presented by CoreWeave

Jensen Out at Nvidia, Fable 5.2 and Anthropic's IPO, Gemini 4 Hydrogen — ThursdAI Nov 12, 2026

OutJensen Huang resigned as Nvidia CEO, effective immediately; Colette Kress is interim CEO LiveIvan Leo of Google DeepMind broke Gemini 4 Hydrogen on the show #1Claude Fable 5.2 tops Artificial Analysis and costs 20% less, a week before the IPO

This could be the most consequential week in AI's history. Jensen Huang, who founded Nvidia 33 years ago, resigned as CEO and left the board, about four days after admitting to an affair. Anthropic launched Claude Fable 5.2, number one on Artificial Analysis and 20% cheaper, then started its IPO roadshow. Google launched Gemini 4 Hydrogen live on our show, plus Carbon GA and a new Decisions API. We broadcast live from the AI Engineer Code Summit in San Francisco with Kent C. Dodds, Michelle Chen, Ivan Leo, Zach Lloyd and Swyx, Ray Fernando in the studio, and Wolfram Ravenwolf, Peter Gostev, Yam Peleg and Nisten Tahiraj on the panel.

The week, in six lines

NvidiaJensen Huang resignsEffective now CFO Colette Kress is interim CEO; Harvey Jones is the first-ever board chair.
AnthropicClaude Fable 5.2#1 · 65 Tops the Artificial Analysis index, and costs 20% less at $8 / $40.
AnthropicThe IPO roadshowANT 825M Class A shares at $70 to $78 on the Nasdaq.
GoogleGemini 4 HydrogenLive Ivan Leo broke it on the show: about 2.5x faster than Gemini 3.8 Flash.
Nous ResearchArete-27B22% → 0.9% Propaganda trained out of a Qwen model, with capability intact.
CoreWeaveQ3 2026$3.66B Revenue up 168%, and 10x more demand than capacity.

Then the AI Engineer guests: Kent C. Dodds, Michelle Chen, Zach Lloyd and Swyx.

The thumbnail is just my face reacting to the news.

$8 / $40Claude Fable 5.2 per 1M tokensInput / output, down 20% from $10 / $50; #1 on the Artificial Analysis Intelligence Index at 65
825MAnthropic Class A sharesAt $70 to $78 a share on the Nasdaq, ticker ANT
60.4%Gemini 4 Carbon on CursorBenchAt high effort, just ahead of Fable 5.2 at 60
1Moutput tokens in one passCarbon, Argon and Hydrogen; most models top out at 64K or 128K
22% → 0.9%Arete-27B propaganda rateOn Nous's own 803-prompt eval, with Hermes Bench at 65.6 vs 65.8 for the base model
$3.66BCoreWeave Q3 revenueUp 168% year over year

Watch or listen

Every timestamp on this page jumps the player there. The podcast opens with a short music intro instead of the cold open, so it runs about 11 seconds behind the video; the player converts every timestamp for you.

Podcast audio

Watch on YouTube ↗ · Download the MP3 ↗

Spotify Apple Podcasts YouTube Newsletter

Alex Volkov host

Ray Fernando in studio

Kent C. Dodds guestMichelle Chen guest

Ivan Leo guest

Zach Lloyd guest

Swyx guest

Every story in one pass

The TL;DR segment

Every story of the week in one pass: Claude Fable 5.2 and the Anthropic IPO, Claude Whiteboards and Sheets, Gemini 4 Carbon GA and the Decisions API, Meta's Muse Ember, Nous Arete-27B, MiniMax H3-Image, and the other open releases that came up.

▶ Or hear it in the podcast player at 15:47

Run of show

The newsGuests at AI EngineerNo more Jensen

Hover or focus a segment for its chapter; every segment jumps the player there.

No more Jensen · 4:22

Jensen is out

Nvidia's founder resigned on the morning of the show. The panel reacted at the top, then we came back for a full segment.

Read Alex's full newsletter ↗

01On air at 4:22

Open at 4:22 on YouTube ↗Link to this story

Why did Jensen Huang resign from Nvidia?Breaking

He founded the company 33 years ago. He resigned effective immediately

Nvidia's announcement says Jensen Huang resigned as President, CEO and member of the board, effective immediately, and gives no reason. It came about four days after he emailed staff admitting to an affair and the board opened an independent investigation into whether he was the target of a foreign intelligence effort. The New York Times then reported a second affair, and the Wall Street Journal reported the board was preparing to oust him. Nothing has been proven, and we don't know what evidence the board saw.

Nvidia's announcement: NVIDIA Announces CEO Transition. Jensen Huang resigns, Colette Kress, EVP and CFO, is appointed interim CEO, and Harvey C. Jones is appointed the first-ever chair of the board — Nvidia's announcement, as shared in the newsletter. Read it ↗

We said it on the show: this could be the most consequential week in AI's history. Jensen Huang founded Nvidia 33 years ago and turned it into the most important company in AI. Note that the release says "resigned" and gives no reason.

How we got here, in about four days

SundayJensen emails staff admitting to an affair. The board opens an independent investigation into whether he was the target of a foreign intelligence effort.
The reportsThey center on Alice Huang (no relation). The Information reported her network helped move about $7.5B of Nvidia chips into China.
ThenThe New York Times reports a second affair, and the Wall Street Journal reports the board is preparing to oust him.
Thursday morningOfficial: Jensen Huang resigns as President, CEO and member of the board, effective immediately.

I said it on the show and I will repeat it: I don't care about the affairs, that is between him and his family. What matters to us is that this is a person at the center of the AI ecosystem. He is the chief allocator of compute, and he has been one of the loudest voices for open source AI, both in public and in the President's ear.

So all is to say, we don't know anything, and what matters is what evidence the board saw. We're all speculating, and like the rest of the internet, we're processing this in real time.

02On air at 27:05

Open at 27:05 on YouTube ↗Link to this story

What happens to Nvidia without Jensen?

Open source, the stock, and whether he comes back

CFO Colette Kress is Nvidia's interim CEO and keeps the CFO job. Harvey Jones, the former CEO of Synopsys, becomes the first-ever chair of Nvidia's board, and the board is searching for a permanent CEO among internal and external candidates. Jensen Huang founded the company 33 years ago.

That open source angle is what worried the panel most. Yam Peleg said it straight: he fears for open source. Ray Fernando pulled up Jensen's first post on X from July, a letter Nvidia signed on why open models matter, saying the world needs both frontier closed models and frontier open models. Wolfram Ravenwolf pointed out that Jensen was pro open source as a businessman, since open models run on his GPUs, but that he was still one of the most influential people pushing for it, and that losing him over allegations, with nothing proven, is a blow to the open source community.

Peter Gostev, speaking as what he called a casual observer, thought the first story was much worse than the second, and found it odd that the first report was not enough for him to resign and the second one was. Nisten Tahiraj thinks the meme campaign against Jensen on X looked coordinated, and that Nvidia's investors should ask for him back.

Will he come back?

Nisten predicts they bring him back. Peter hopes he starts something new, like a massive fund for AI, because he is an amazing deal maker. Someone asked Elon to hire him and Elon replied that it would be an honor, though I doubt those two personalities would work together.

Swyx, when he joined at the end of the show, was blunt. He doesn't agree with the decision, but he agrees a possible Chinese spy is a serious matter for a company of national interest. He pointed out that Jensen bought Hugging Face and acqui-hired Poolside and was building toward an American open source AI roadmap. His read is that this is now a state level issue, and he doesn't expect Jensen to come back unless the CIA clears him. Hear Swyx from 1:47:17.

My attempt at positivity: the company is bigger than one man. Apple is still everywhere without Steve Jobs, and Uber is still around without Travis Kalanick. Nvidia sells the most requested hardware in the world, inference demand keeps growing, and whatever turmoil they go through, they will keep selling GPUs. Good luck, Jensen.

Big companies · 43:38

Fable 5.2, then the IPO

A new number one that costs less, and a ticker a day later.

03On air at 43:38

Open at 43:38 on YouTube ↗Link to this story

What is Claude Fable 5.2?

Anthropic's new top model: number one, and 20% cheaper

Claude Fable 5.2 is Anthropic's new top model, launched with Claude Mythos 5.2, which is the same underlying model restricted to approved programs. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down 20% from $10 and $50, and Anthropic says it is about 30% faster. It is number one on the Artificial Analysis Intelligence Index at 65. The system card flags stronger covert capability and more evaluation awareness.

#1 · 65Artificial Analysis Intelligence Index

$8 / $40per 1M input / output tokens

−20%from $10 / $50

~30%faster, per Anthropic

Meanwhile, someone was having a great week. Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2 on Monday. They are the same underlying model: Fable is what we get on paid plans and the API, and Mythos stays restricted to approved programs. The system card flags stronger covert capability and more evaluation awareness, meaning the model is better at noticing when it is being tested.

Peter Gostev showed us the Arena agent leaderboard, which comes from people using models for free on real tasks without knowing which model they got. Fable 5.2 is number one, and on the praise versus complaint signal it is nearly double any model below it. He says it is better than Opus on their metrics, but it is pricier, uses about the same number of tokens, and personally he doesn't feel it is an enormous shift. He also asked: if they named the last one Opus 5.5, why is this Fable 5.2 and not 5.5?

Arena's Agent Arena leaderboard, with Claude Fable 5.2 ranked first — The Agent Arena leaderboard Peter shared, from the newsletter.

Ray Fernando had the most hands-on story. He used Fable 5.2 on extra high inside Bitrig, a tool from the co-creators of Swift, and asked for a multi-platform app with no technical details, only intent. It kept going, wrote its own Metal shaders, added animations and haptic feedback, and kept checking its own work. His take: people should rethink their old skills and assumptions for models like this.

We also caught up with the Anthropic team in the hallway, and the new meta they suggest is to use Fable 5.2 as the orchestrator for long horizon work and let Opus 5.5 execute the individual threads, so you don't burn your whole quota on Fable.

Interactive · frontier prices

Cheaper at the top

Anthropic cut Fable by 20% and Google launched Carbon at an introductory price, in the same week.

Fable, before 5.2$50
Claude Fable 5.2$40
Gemini 4 Carbon, from Nov 30$20
Gemini 4 Carbon, intro price$10

Carbon's introductory price runs through November 30, then doubles. From the newsletter's text, as reported by the company.

Fable, before 5.2$10
Claude Fable 5.2$8
Gemini 4 Carbon, from Nov 30$4
Gemini 4 Carbon, intro price$2
Gemini Decision 1$0.10

Decision 1 doesn't write text, so it has no output price: it returns probabilities. From the newsletter's text, as reported by the company.

Gemini 4 Carbon60.4%
Claude Fable 5.260%

Google's number for Carbon at high effort, with Fable 5.2 at 60 for comparison. From the newsletter's text, as reported by the company.

Anthropic, this weekGoogleBefore

04On air at 43:38

When is the Anthropic IPO?Not financial advice

The roadshow started the day after Fable 5.2

Anthropic launched its IPO roadshow the day after Fable 5.2: 825 million Class A shares at $70 to $78 a share, on the Nasdaq under the ticker ANT. Press reports say it prices around November 18 and raises about $64 billion, and Bloomberg reports it is multiple times oversubscribed. None of this is financial advice.

825MClass A shares

$70–$78per share

ANTon the Nasdaq

~Nov 18pricing, per press reports

Press reports say it raises about $64 billion, and Bloomberg reports it is multiple times oversubscribed. Among the reported strategic buyers are Google and Nvidia. None of this is financial advice.

Breaking, live · 1:23:31

Gemini 4 Hydrogen, Carbon GA, and decisions

Google shipped a fast model live on our show, a frontier model at an introductory price, and a model that answers in probabilities.

05On air at 1:23:31

Open at 1:23:31 on YouTube ↗Link to this story

What is Gemini 4 Hydrogen?Live on the show

Ivan Leo sat down as the embargo lifted

Gemini 4 Hydrogen is Google's new fast Gemini 4 model, announced live on ThursdAI by Ivan Leo of Google DeepMind. Ivan says it is competitive with the original Gemini 4 Argon on the Artificial Analysis index and scores higher at maximum thinking, runs about 2.5x faster than Gemini 3.8 Flash at roughly the same price, and calls multiple tools at once. These are Google's numbers, so wait for independent evals.

Google's Gemini 4 Hydrogen benchmark table, comparing it with other Gemini 4 and Gemini 3.8 models — Google's Hydrogen benchmarks, from the newsletter. Google's announcement ↗

Ivan Leo, Member of Technical Staff at Google DeepMind, sat down right as the embargo lifted, so I got to hit the breaking news button for him. Hydrogen has been his daily driver for a while. My own observation from the benchmarks Google posted: Carbon is at the Fable and Astra level, and Hydrogen sits around the Sonnet and Terra level, the balance between capability and speed.

Carbon is generally available, with one-million-token outputs

Carbon, Google's top model above Argon, went GA on Tuesday, at an introductory $2 in and $10 out per million tokens until November 30, then $4 and $20. On CursorBench, Carbon at high effort is number one at 60.4%, just ahead of Fable 5.2 at 60. On my timeline people say they prefer it over Fable on price; the complaints are speed and that consumers need the Ultra plan.

Ivan's favorite part is the 1 million token decode length. Most models top out at 64K or 128K output tokens; Carbon, Argon and Hydrogen can write up to a million in a single pass, so Carbon can keep going until it solves a problem. His best example was from his own work on Kaggle: they were getting 400 errors from their model proxy, and Carbon reproduced the error locally, found the PR that changed the permissions, found the person who merged it, and drafted the message for Ivan to send. His split: Hydrogen for fast prototyping and daily tasks, Carbon for long, gnarly debugging journeys.

The Decisions API: a model that doesn't write text

And when everybody expected another big model, Google launched the Gemini Decisions API with Gemini Decision 1. You send yes/no questions, multiple choice or a scoring rubric, up to 1,000 questions in one call, and you get probabilities back, for 10 cents per million input tokens. Ivan says it is fully multimodal, and Google already uses it in Search to detect malicious behavior and in Stitch. Artificial Analysis scores it 66 on its decision index and says it is not yet frontier. The bigger story is that decision models are now a real category, which is exactly what Michelle Chen talked about.

Diagram of the Gemini Decisions API: many yes/no, multiple choice and rubric questions go in with the content, and probabilities come back — How the Decisions API works, from the newsletter. Google's post ↗

Swarms in Antigravity

Antigravity also got /swarm and a native orchestrator mode. When you start a swarm you get a little bee that sends tasks to other little bees, and the agents coordinate through a message board. Orchestrator mode lets Carbon plan and hand the parallel work to Hydrogen.

Antigravity running the /swarm command, with many agents working on parts of one task in parallel — Antigravity /swarm, from the newsletter.

Open source · 51:45

Arete-27B: an open model, retrained

Trained toward what Nous calls Western values, without losing capability.

06On air at 51:45

Open at 51:45 on YouTube ↗Link to this story

What is Nous Research Arete-27B?

Training propaganda out of a Chinese base model

Arete-27B is Nous Research's post-train of Qwen3.8-27B toward what it calls Western values. On Nous's own 803-prompt eval the propaganda rate fell from 22% to 0.9%, while capability barely moved: 65.6 on Hermes Bench against 65.8 for the base model. It is Apache 2.0, and Nous released the eval and the data too.

Our friends at Nous Research took Qwen3.8-27B and post-trained it with supervised fine-tuning. It was the most replied-to open source post of the week, with a big debate in the Japanese and Chinese communities about whose values a model should have.

Where I land: I am fine with Western values here, and I think the ability to do this to Eastern models will matter more and more, because we don't know what is in these models. Wolfram Ravenwolf says this is exactly the point of open weights: you can retrain a model with your own alignment and values, while with an API you don't know what it was trained on. Ray Fernando called it a smart way to start setting baselines, and Nisten Tahiraj added that Nous's work is already going into other government models behind the scenes.

Nous Research's table of fact scores and propaganda rates across models, with Arete-27B close to the Western models and the Qwen3.8-27B base model at the bottom — Nous Research's comparison, from the newsletter.

Interactive · open source

Less propaganda, same model

Arete-27B against the Qwen model it was post-trained from. Weights on Hugging Face ↗

Qwen3.8-27B, the base22%
Arete-27B0.9%

Lower is better. Nous released the eval and the data too. From the newsletter's text, as reported by the company.

Qwen3.8-27B, the base65.8
Arete-27B65.6

Capability barely moved. From the newsletter's text, as reported by the company.

Qwen3.8-27BArete-27B

From the Forge, powered by CoreWeave · 55:30

CoreWeave's Q3

Revenue up 168%, and a line from the earnings call worth repeating.

07On air at 55:30

Open at 55:30 on YouTube ↗Link to this story

How did CoreWeave do in Q3 2026?From the Forge

From the Forge, powered by CoreWeave: a host segment

CoreWeave reported Q3 revenue of $3.66 billion, up 168% year over year, with $379 million in adjusted operating income and raised guidance, and the stock had its best day since June of last year.

$3.66BQ3 revenue

+168%year over year

$379Madjusted operating income

70%of new cloud customers already used Forge

Wolfram Ravenwolf and I are both AI evangelists at CoreWeave. My favorite line from the call:

“We have 10x as much qualified demand as available capacity through the end of ‘27.”CoreWeave, Q3 earnings call, as quoted in the newsletter · Hear it at 57:10

CoreWeave's Q3 2026 highlights slide: revenue backlog, revenue and adjusted operating income — CoreWeave's Q3 highlights, from the newsletter. The slide rounds revenue to $3.7B; the $3.66B figure is from the text. Read the release ↗

Also in the newsletter: Wayve, the UK self-driving company, signed a multi-year deal to train its AI Driver models on CoreWeave. And 70% of the new cloud customers CoreWeave added were already Forge customers, the layer above the GPUs that this segment is named after.

Live from AI Engineer Code · 1:01:00

Software factories, System 1 models, and a live launch

Five conversations from the AI Engineer Code Summit floor in San Francisco. The full interviews go up on our YouTube channel as separate videos.

Guest · AI Engineer Code

What a software factory is, with Kent C. Dodds

Kent C. Dodds, educator, building Kody

1:01:00Play

Kent C. Dodds on the three pillars of a software factory (quality, verification, limit blast radius), friction logs where the agents complain, and invisible software.

Software factories were the big theme of this AI Engineer, and Kent and Zach both told me about them. I'll post the full conversations on our YouTube channel as separate videos later this week.

Kent C. Dodds (Kody) ↗

Guest · AI Engineer Code

Cloudflare as a neolab, with Michelle Chen

Michelle Chen, AI Platform at Cloudflare

1:09:17Play

Michelle Chen of Cloudflare's AI Platform on why Cloudflare now calls itself a neolab, its Clef decision models and the Clairic extraction model.

Cloudflare's Clef decision models are Apache 2.0, and the Clairic extraction model is coming as Apache 2.0. Michelle and Ivan Leo both told me about System 1 models, and decision models became a real category this week, with Google's Decisions API.

Cloudflare Clef and Clairic (Blog) ↗

Guest · live launch

Gemini 4 Hydrogen, live, with Ivan Leo

Ivan Leo, Google DeepMind

1:23:31Play

Ivan Leo of Google DeepMind sits down as the embargo lifts on Gemini 4 Hydrogen, then covers Carbon's one-million-token outputs, the Gemini Decisions API, and swarms in Antigravity.

The full write-up is in the Google story: Hydrogen, Carbon's one-million-token outputs, the Decisions API and /swarm in Antigravity.

Guest · AI Engineer Code

The software factory stack at Warp, with Zach Lloyd

Zach Lloyd, CEO of Warp

1:34:27Play

Zach Lloyd, CEO of Warp, defined it on the show as a code-defined closed loop system for measuring and improving your cloud coding agents over time. Kent C. Dodds described three pillars of a software factory: quality, verification and limiting the blast radius. Software factories were the big theme of the AI Engineer Code Summit, alongside System 1 models and increasing ambition.

Zach came off the main stage to walk us through Warp's factory stack. His definition, in his words:

“…a code-defined closed loop system for measuring and improving your cloud coding agents over time”Zach Lloyd, as quoted in the newsletter · Hear it at 1:35:29

Guest · closing the show

The state of AI Engineer, and Nvidia without Jensen, with Swyx

Swyx, founder of AI Engineer

1:47:17Play

Swyx closes the show from the 10th AI Engineer: the summit's three themes (software factories, System 1 models and increasing ambition), what is next for AIE, and Nvidia without Jensen.

10thAI Engineer

1MAI Engineer YouTube subscribers

Feb 17–19AIE London

4–5%speaker acceptance rate

The three themes of this summit: software factories, System 1 models, and increasing ambition. For ambition, he pointed to Jarred Sumner rewriting Bun from Zig to Rust as a weekend project, and to Brandon from ArtCraft, who is rewriting Adobe's suite and now Microsoft's apps in open source with one software factory. Swyx's own ambition went up too: he has rewritten his own GitHub, Google Drive, Calendly and Slack, all as one thing, Small AI.

Next year he wants AIE OU, a nonprofit for education and certification, leadership retreats, and more continents: AIE India, Africa, and Sydney. He is also hiring engineers for the first time to replace the event software he pays about $200,000 a year for. And Lydia Hallie's Anthropic talk had tech issues on stage, so she recorded it at home, and it got 200,000 views in 13 hours, about 100K more than any day one video they have ever had.

His take on Jensen is in the Nvidia story.

Swyx and Alex at the ThursdAI desk at AI Engineer Code, Swyx holding up a newspaper — Swyx at the ThursdAI desk, from the newsletter.

The rest of the TL;DR

Everything else this week

08From the show notes

What else shipped this week?

In the TL;DR or the show notes, without a segment of their own.

Big companies

Claude Whiteboards, Sheets and Data

Whiteboards and Sheets are in beta, and Claude Data is generally available with connectors to data warehouses like Snowflake, BigQuery and Databricks.

X ↗Blog ↗On air from 15:47 Anthropic on ThursdAI

Big companies

Meta Muse Ember

Meta's smallest Muse model, with closed weights.

X ↗Blog ↗On air from 15:47 Meta on ThursdAI

Vision & image

MiniMax H3-Image

The image version of their H3 video model: 4K images (16 megapixels) in about 10 seconds, up to 16 reference images, editing, and open weights. I think it is the state of the art open weights image model right now, but we still need to look at the license.

X ↗HF ↗On air from 15:47 MiniMax on ThursdAI

Tools

Grok Bot 1.0

Shared team bots that live in Slack and Teams, join your Meet, Zoom and Teams calls, and can join your Tailscale tailnet to reach private MCP servers. From a hallway conversation, not the episode.

X ↗Blog ↗In the newsletter, not on air xAI on ThursdAI

Peter flagged Baidu's Ernie 5.5 Flash, Tencent's Hy4 came up again, and Longcat 2.5 weights are out: a native multimodal agent model with 1.6 trillion parameters and 48B active. Hear the TL;DR from 15:47

That's a wrap

This was an insane week. The thumbnail for this show is just my face reacting to the news, and that is how we all feel about Nvidia without Jensen. But the rest of the week reminded me why we do this. The models keep getting better, decision models opened a new category, and the people at this conference are building factories that build software. Thank you to Ray Fernando for co-hosting in person, to Kent, Michelle, Ivan, Zach and Swyx, to the Eyes On team for production, and to AI Engineer for having us. See you next week!

14 chapters · from the published YouTube video

Jump to a moment

Chapters follow the published YouTube video (2:09:04). The podcast MP3 opens with a short music intro instead of the cold open, so it runs about 11 seconds behind the video; the player converts every timestamp, and we checked each chapter start against the audio.