What's Actually Happening
Anthropic is a public company. The Claude maker started trading on Nasdaq today as ANT after raising about $81.3 billion at $83 a share, the largest IPO on record, and the stock closed its first day at $97.17, up 17%.
The day's other big story is one you can use tonight: Microsoft's in-house MAI Code 2 is live in GitHub Copilot, claiming near-frontier coding for a fraction of the price, with a Flash version on Copilot's free plan.
Also tonight: Google opens Genie 4 worlds you can rewrite with a sentence, Cloudflare open-sources a model that turns messy documents into clean data, and four tools worth trying.
MARKETS
📈 Anthropic Goes Public As ANT
Anthropic priced 980 million shares at $83 on November 18 and began trading on the Nasdaq Global Select Market today under the ticker ANT. That's about $81.3 billion raised, topping SpaceX's $75 billion June listing as the largest IPO ever, according to Forbes.
The first trade came at $111.50, about 34% above the offer price. Shares peaked near $115, slid as low as $92.72, and closed at $97.17, up 17.1%, for a market value of roughly $1.27 trillion. The prospectus shows both why investors are paying up and why they're nervous: revenue hit $16.3 billion in the first half of 2026, up from $1.2 billion a year earlier, alongside a net loss of $80.7 billion. Amazon owns nearly 11% and Alphabet nearly 10% on a fully diluted basis.
Our take: a 34% pop that cooled to 17% by the close says the market loves Claude's growth but hasn't made peace with what it costs to build. Nothing changes for Claude users overnight, but Anthropic now answers to shareholders every quarter, and its pricing, usage limits and launches will be read through that lens. Watch the first earnings call.
Go deeper: Forbes' day-one recap · ANT's day, trade by trade
NEW MODEL
⚡ Microsoft Drops MAI Code 2 In GitHub Copilot
Microsoft announced MAI Code 2 and MAI Code 2 Flash at Ignite on November 17, and both are already in the GitHub Copilot model picker: VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, the cloud agent, github.com and GitHub Mobile. MAI Code 2 is on Copilot Pro, Pro+, Max, Business and Enterprise. Flash is on every plan, including Free and Student.
The pitch is cost per result. Averaged across four coding benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's at 5% of the cost. In a blind frontend test with 258 annotators, developers preferred its results over GPT-6 Sol 64% of the time. Flash, built for quick edits, explanations and routine fixes, matches GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their cost. Both are billed at a 50% promotional price for now; The Decoder reports MAI Code 2 at $0.50 input and $2.50 output per million tokens in Copilot through May 17, 2027.
Our take: Microsoft is upfront about the limits: it says neither model is tuned yet for very long-horizon agentic work or fundamental research, and every number above is its own. But most daily coding is multi-file fixes, debugging and UI work, and a model that's nearly as good for a fraction of the price is a real reason to flip the model picker. If you're on Copilot Free, Flash costs you nothing to test tonight.
Will you try MAI Code 2 in GitHub Copilot this week?
🧩 Why It Matters: Microsoft Builds Its Own Stack
MAI Code 2 isn't a one-off. Microsoft AI shipped seven models at Ignite, including MAI Cyber 2 Flash for security work, MAI Image 2.7, which Microsoft says ranks #1 for image editing on Artificial Analysis ahead of GPT Image 2.6 Sunburst at one-fifth the price, MAI Realtime 1, its first live speech-to-speech model at one cent per audio minute, and Decision 1.1 for routing. MAI Work Flash, tuned from Code 2 Flash, already runs across Microsoft 365, where Microsoft says it matches GPT-6 Luna on Excel and PowerPoint tasks.
Notice who every chart is measured against: OpenAI and Anthropic, the partners whose models fill Microsoft's products. Microsoft hasn't said it's dropping them, and Copilot still offers their models. But cheap in-house models give it pricing leverage and a fallback if those partnerships shift. For everyone else, it puts a third deep-pocketed lab into the coding price war, the same week Google sold Gemini 4 Neon on cost per task. Expect the price of near-frontier coding to keep falling.
A Word From Today's Sponsor
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
Let’s Talk MODELS! 🤖
Google released Genie 4 in Project Genie on November 18, rolling out to eligible Google AI Ultra subscribers, with a waitlist for AI Pro members. Genie generates interactive worlds you can move through, and Genie 4 lets you type a new prompt mid-session to change the weather, add a character or redirect the action. It's also the first Genie model with real-time native audio, from ambient soundscapes to character dialogue.
Genie 3, still in the model picker, now runs sessions up to three minutes instead of 60 seconds, and worlds can be chained into longer journeys or seeded from 360-degree panoramas. A new agent builds a world's opening visuals and a custom soundtrack from a plain-language description, and a community gallery lets you publish worlds and remix other people's, down to their rules and music. All generated video and audio carries SynthID watermarks.
Our take: Google still calls this frontier research and lists the rough edges itself: camera latency, tricky physics and scenes that drift over time. But world models are moving from demo reels to something paying users can actually play with, and rewriting a world with a sentence is a genuinely new kind of interface. If you have Ultra, give it ten minutes tonight.
Try it: Project Genie · Genie 4 research overview
The Model That Reads Your Receipts
Cloudflare released open weights for Clairic (27B parameters) and the new Clairic-flash (9B) on November 18. They're extraction models: hand them text, images or PDFs and they return typed values, like a receipt's merchant, line items and total, with confidence scores and the source text behind each answer. They also handle estimates, forecasts and date math. Clairic is built on Qwen3.8-27B and Clairic-flash on Qwen3.5-9B; Unite.AI reports an Apache-2.0 license, so check the model card before shipping.
On Workers AI, both cost $0.10 per million input tokens until December 1, with no charge for output tokens; after that, Clairic goes back to $0.24. Cloudflare's own tests show near-perfect field accuracy and a median 155 ms for Clairic-flash on text fields, though it notes there's no public extraction benchmark to compare against.
Our take: pulling fields out of documents is one of the most common, least glamorous jobs in AI apps, and most teams overpay by sending it to a general chatbot. A small open model that returns structured fields with confidence scores, and that you can run yourself, is the right tool for it.
Try it: Clairic weights · Clairic-flash weights
Top 5 In AI Research 🔬
The other stories worth your attention:
Kuaishou launched Kling Image 4.0 with native 8K generation, its biggest image-model upgrade yet as Chinese labs push on resolution. See what else changed in the new version. (TechNode)
OpenAI says its models are now redesigning chips, including parts of its own Jalapeño inference chip and blocks of AMD's flagship GPU designs, and it's training a new model with Synopsys. See how big the gains were. (OpenAI)
OpenAI measured how often its agents finish multi-hour research tasks alone, and the success rate has climbed sharply since March. See which jobs agents ace and which they still fail. (OpenAI)
More big US banks say AI-agent losses are on you, even if the agent gets hacked, as Citi, PNC and Citizens follow JPMorgan and Bank of America. See the one exception Citi carved out. (TechCrunch)
Palo Alto Networks is buying Applied Compute, a platform that helps companies post-train and run open-weight models, in an all-stock deal. See what it's paying and who becomes its co-chief AI officer. (Palo Alto Networks)
🛠️ Tools That Are Hot Right Now!
👻 Snap Easy Lens CLI - build, test and publish Snapchat Lenses by prompting Claude Code, Codex or Cursor from your terminal. Free on Mac, Windows and Linux; see what it takes to go from prompt to a live Lens.
🧠 Perplexity Brain skills - Perplexity's Computer agent now writes and updates its own reusable skills from your past tasks, so repeat jobs get smoother. On by default for US Max subscribers (paid); see how much it cut failed tool calls.
🧾 ChatGPT for Accounting - reconcile transactions, prep invoice entries and draft month-end reports from your spreadsheets and accounting apps. Free to install on ChatGPT Business and Enterprise (not the free plan); see which apps it connects to.
🎨 MAI Labs Canvas - make an image, then edit any part of it, in Microsoft's experimental image workspace powered by MAI Image 2.7 Flash. A limited preview with no price listed on the page; see what else you can try in MAI Labs.
What's The Recap?
Anthropic went public as ANT in the largest IPO on record, raising about $81.3 billion at $83 a share and closing day one at $97.17, up 17%. Microsoft put MAI Code 2 in GitHub Copilot, claiming 98% of GPT-6.1 Sol's coding performance at 40% of the cost, with a Flash version free for every Copilot user. Google opened Genie 4 to Ultra subscribers with worlds you can rewrite mid-session, Cloudflare released open Clairic models for pulling clean data out of documents, and banks told customers that what their AI agents do is on them. The thread: AI is now big enough to be priced like the economy's next core layer, and cheap enough to use everywhere.
What Do You Rate Today's Newsletter?
Stay building. 🤖
We put Gemini 4 Neon through a full test against Opus 5.5 and GPT-6.2.
📬 Read by 14K AI readers. Know someone who'd like it? Send them here to subscribe.



