This website uses cookies

Read our Privacy policy and Terms of use for more information.

Together with

Date: Sunday, November 22, 2026

Hey {{first_name | AI enthusiast}},

Anthropic's market debut and Nvidia's leadership change put two AI giants under a new kind of scrutiny. Meanwhile, Factory, Stripe and OpenAI show how the model race reaches everyday work.

Also: Did OpenAI just launch one of its worst models?

In this edition

Best regards,

PS: If you want to organize corporate AI training for your employees »

AI MARKETS

Anthropic's record IPO is only the opening act

Anthropic sells shares at $83 and raises $81.3 billion in the largest IPO on record. The offer values the Claude maker at about $1.5 trillion on a fully diluted basis. Its shares open at $111.50, then close their first day at $97.17, up 17% from the offer price.

That debut is a striking vote of confidence, but the price also moves sharply within a single day. Chamath's analysis and The Rundown's coverage point to the same question: can Anthropic turn rapid adoption into durable returns while paying for immense computing capacity? The answer will emerge in future results, not the first trading session.

Anthropic’s most recent disclosed balance sheet (September 30, 2026) showed $95.6 billion in cash and cash equivalents. Including $36.2 billion in short-term investments, its total cash-and-investments position was about $131.8 billion before IPO proceeds. That is one helluva cash chest!

So what? For founders and investors, revenue growth deserves attention alongside the cost of serving each customer and the cash needed to keep growing.

LEADERSHIP

Nvidia moves on from Jensen Huang

Jensen Huang resigned as Nvidia's chief executive and director after revelations of an alleged affair with a Chinese operative. CFO Colette Kress has taken the interim CEO role, and Harvey Jones has become the board chair. A public push for Huang's return followed, but Jones says the board does not intend to reinstate him.

The company must now show that its product roadmap and customer relationships can survive a founder's abrupt exit. Strong demand for AI chips makes the succession especially consequential. The board's next permanent appointment will shape how Nvidia navigates customers, regulators and its own fast expansion.

NVIDIA’s own disclosed order book is at least $1 trillion in Blackwell and Vera Rubin orders through the end of 2027, announced by CEO Jensen Huang at GTC in March 2026. The Jensen episode itself caused an immediate roughly 13% weekly hit and about $1.7 trillion of market-cap loss during the associated losing streak; the much larger 46%–48% decline reflects the episode combined with wider AI, regulatory, and accounting concerns.

Markets dropped billions because the perception is that the entire AI trade is crumbling. Overreaction or reality- we will know in a few months.

So what? Founder-led companies need succession plans that keep decisions, customer trust and execution moving when a leader suddenly leaves. Now that strange way of having no 1:1s or a large span of control does not look very appealing.

AI ECONOMICS

Factory Router lets the task choose the model

Factory Router is a tool that I chanced upon recently and it selects a model for each coding session automatically. It sends simpler work to cheaper models and keeps demanding work on stronger ones. If the first choice struggles, it can switch to a more capable model.

Factory says routing cuts inference costs by 63% across its production sessions. That is the company's reported result, measured against using frontier-model rates for the same work. It is not a promise that every team will save 63%. But it is definitely worth trying.

So what? Try routing a sample of repeatable tasks, then compare completed work, error rates and total cost against your current model setup. Could be pivotal for startups!

PRACTICAL AI

Stripe finds the gap between a good score and a working integration

Stripe gives coding agents 69 realistic payment tasks, from checkout to subscriptions and marketplace payments. Across 14 agent setups, the average score is 78.1%. But only 29.4% pass every essential requirement.

That gap matters. An agent can produce a convincing checkout while missing how to handle a failed payment or a subscription change. Stripe also finds that its guidance improves every agent setup tested, by roughly three to eight percentage points.

There is a limit to the result: Stripe checks the agents’ work against detailed criteria, but does not run the finished integrations. A passing score is not proof that the code works in production.

So what? When you assess AI work, list the steps that must succeed. Test failures as well as the happy path, and require evidence for every essential step. An impressive average can hide the one miss that matters.

MODEL RACE

OpenAI's new models need a test on your own work

OpenAI has introduced GPT-6.2 Astra for harder projects and GPT-6.2 Sol for everyday work. It says the models improve professional documents, coding and multistep tasks while using fewer output tokens. Those performance comparisons come from OpenAI's own release materials.

The Rundown newsletter notes a mixed early reception. That is a useful reminder to check the work itself: a spreadsheet should reconcile, a slide should make its point and code should run. A benchmark lead or a launch claim alone cannot tell a team which model fits its process.

The reaction on X to OpenAI’s latest releases—GPT-6.2 Astra and GPT-6.2 Sol—has been sharply negative among prominent AI developers and power users, despite acknowledgment that the models are highly capable. The main complaint is not raw intelligence but poor usability, verbosity, and difficulty steering on real work.

What X users are saying

  • Matt Shumer, investor and former HyperWrite CEO, said he had “a terrible experience testing them,” calling them “essentially unusable” and saying he would not recommend them

  • Theo Browne, CEO of T3 Chat, called GPT-6.2 Astra “the worst model OpenAI has ever released,” saying it repeats itself, overcomplicates small tasks, and is hard to use on long projects.

  • Ethan Mollick, Wharton professor, called it the first OpenAI frontier release since GPT-4.5 that was not a clear step forward; he said it could do impressive work but was often frustrating and difficult to guide.

So what? Test the same three recurring jobs across your current model and the new release, then compare accuracy, editing time and cost.

AI is only as good as what you tell it

The fastest way to give AI more context is to talk. But that only works if the transcription is right. Wispr Flow runs on Canto, its own speech model, trained on real dictation instead of studio audio.

In Wispr's tests on 10 hours of real dictations from 2,300+ speakers, Canto had the lowest word error rate of every model tested, including Google and OpenAI. It also pulls from your personal dictionary, so tool names, project names and teammates come out spelled right.

Talk through the whole problem, hit enter, and get a better answer. 4x faster than typing, in every app.

Try this: make your AI work pass every requirement

Stripe's strict pass rate suggests a simple habit: ask for a requirement-by-requirement review before you accept an AI output.

Act as a strict reviewer of this AI-generated deliverable. 
First list every requirement in my brief. 
Then check each requirement against the output and mark it Pass,
Fail or Unclear. Test edge cases, failure paths and missing 
steps. Quote the exact evidence for each Pass. 
Do not award an overall Pass unless every required item passes. 
Finish with the smallest set of fixes needed and a test for each 
fix.

My brief: [paste the task and acceptance criteria]

AI output: [paste the result]

Replace the brackets with a real task. Check the review yourself before relying on the result.

Hiring abroad without an entity? Remote becomes your hire’s legal employer, with an average onboarding time of 3 days. We handle contracts, benefits and payroll taxes in line with local requirements.

State of Product reveals what AI still hasn’t solved

80% of product professionals say AI helps them ship faster, but customers aren’t seeing value any sooner. Atlassian’s State of Product 2027 explores the gains AI is delivering and the challenges that remain, from decision-making that hasn’t kept pace to gut instinct overriding customer evidence.

How did you like this edition?

Your feedback helps us to improve.

Login or Subscribe to participate