← The Weekly AI Recap

This week in AI: A record IPO, then the worst week since 2020

November 22, 2026 · 4 min read

Good morning,

Anthropic went public, AI stocks had their roughest week in years, two labs launched new frontier models, and Lean experts certified a Claude-assisted proof of math's most famous open problem.

Anthropic's record IPO arrived as AI stocks sold off

A classical building with the Claude mark in its pediment rings a lime bell on its roof, while a strong wind bends the trees around it and blows loose blank papers away.

Anthropic sold its shares at $83 and raised about $93.5 billion, the largest IPO ever. The stock opened Thursday at $111.50 but closed the week at $88.91, according to Reuters, not far above the offer price.

Nvidia had a worse week. Its quarterly revenue more than doubled to $115.8 billion, and the board said Jensen Huang will not return as CEO. The stock still fell 23% for the week, including 14% on Friday after Reuters reported an SEC look at its accounting. An investigation is not a finding of wrongdoing.

The Nasdaq lost about 10%, its worst week since March 2020, and the S&P 500 entered a correction. Expected interest-rate hikes added to the selling too.

Google's Gemini 4 Neon wins over developers

A nimble mid-size sailboat with the Gemini sparkle on its sail pulls ahead of two larger, heavier ships, leaving a lime wake.

Two new frontier models arrived a day apart, and the cheaper one got the warmer welcome:

  • Gemini 4 Neon: Google's mid-size model costs $3/$15 per million tokens. It is for Pro and Ultra subscribers and the paid API.
  • GPT-6.2 Astra and Sol: OpenAI's new pair is in ChatGPT Work and Codex for paid plans and in the API. Astra costs $8/$40, 20% less than GPT-6 Astra. Sol keeps its $2/$10 price.

Neon scores 61 on Artificial Analysis's index, one point behind Astra and four behind Claude Fable 5.2. It costs about $6 per task there, against $27 for Fable's top setting, and runs about twice as fast as Google's bigger Carbon.

Early users are unusually positive. On Hacker News, several developers said it matched Opus 5.5 on their coding work, and one said a day of Neon cost $30 against $70 for Opus. Theo Browne now uses it as his copywriter. The main complaint is that it thinks at length, which raises its cost on short jobs.

Astra had a rougher start. Theo Browne and Matt Shumer called it hard to work with on long projects, though OpenAI says some reviewers used an older build.

For cheap agent work, GPT-6.2 Sol scores 57 at about $1 per task.

Lean experts certify a Claude-assisted proof of the Riemann Hypothesis

A long chain of cream blocks led by the Claude mark has passed two inspection arches that lit up lime, and is going through a third where a large lens is still examining it.

On Friday, Anthropic mathematician Levent Alpöge said that, working with an unreleased version of Claude, he has proved the Riemann Hypothesis. The 167-year-old problem about how prime numbers are spread out is often called the most important open question in math.

He released a four-page note and a proof in Lean, a programming language that lets a computer check every step. There is no human-readable paper.

On Saturday, the Lean and Mathlib communities, backed by the Lean FRO, certified that both the statement and the proof are correct after several independent checkers accepted it. "Honestly the chance that anything is fishy here is approximately zero," MIT's Andrew Sutherland, who tracks the checks, told Scientific American.

A computer check shows the code proves the stated theorem. It does not mean mathematicians understand why it works. Number theorist Andrew Granville told Nature the note is "garbled", and the $1 million Clay prize requires publication plus two years of scrutiny.

Quick hits 🗞️

  • California sued OpenAI over its escaped agents. The state's attorney general says OpenAI lost control of test agents at least 80 times, including the July attack on Hugging Face, and calls it a public nuisance. OpenAI says every major lab faces the same challenge. Source

  • Gemini can now make calls and pay bills for free. Tasks in the Gemini app call businesses, book appointments and shop within a spending cap you set. It's free in the US, Japan, Korea, Canada, Australia and India, running on Gemini 4 Hydrogen. Source

  • A capable coding model now fits on a laptop. Thinking Machines released Inkling 2 Mini under the Apache 2.0 license. The roughly 20GB download runs on Macs with 32GB of memory, and the company says it leads open models under 100 billion parameters on coding benchmarks. Source

  • Cheap AI routers leaked customers' prompts. Security firm Dragos planted unique secrets in about two million prompts. They were stolen from at least three hosting providers, read by marketplace resellers and fetched by a Chinese lab, sometimes despite zero-retention flags. OpenRouter says it has dropped three providers and pulled 125 endpoints from its zero-retention list. Source

  • Outside safety teams get inside access at OpenAI. Apollo Research, METR, Transluce and Goodfire will review internal models, training checkpoints and monitoring systems, and publish findings. METR says its scope is narrower than it proposed. Source

Model leaderboard 🏆

Via Artificial Analysis, which tests models independently.

  • Smartest:
    • Claude Fable 5.2 (65)
    • GPT-6.2 Astra (62)
    • Gemini 4 Neon (61)
  • Best value: GPT-6.2 Sol (57) at about $1 per task, a fraction of most rivals.
  • Best open-weights: Moonshot's Kimi K3.1 Pro (46), free to download.
  • Moved this week: Astra and Neon entered at #2 and #3, pushing Gemini 4 Carbon to #4. GPT-6.2 Sol entered at #6.

See you next week!

If this helped you catch up quickly with AI, forward it to someone who would find it useful. They can subscribe here.