AI Brief, November 19: Gemini 4 Neon, Anthropic's IPO, GPT-6.2 Astra and Sol
Google ships Gemini 4 Neon, Anthropic prices its IPO above range, OpenAI releases GPT-6.2 Astra and Sol, and Microsoft drops seven MAI models.
By Sean WeldonThree frontier labs shipped or priced something major within a day of each other, led by Google's Gemini 4 Neon, a mid-tier model that benchmarks close to the top of the field. Anthropic's IPO priced above its range ahead of today's Nasdaq debut, while OpenAI and Microsoft both expanded their model lineups with cheaper options for high-volume agent work.
Google ships Gemini 4 Neon, a mid-tier model that benchmarks near the frontier
Google introduced Gemini 4 Neon with a 1 million-token context window and four thinking levels. Google says it ranks #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, #2 on the Vals Index at 70.9%, #1 on the Artificial Analysis Coding Agent Index at 58.6, and #2 on Arena's new Agent Arena Index at 1336. It is rolling out in the Gemini API via AI Studio and Android Studio, Gemini for business, Google Antigravity, and the Gemini app for Pro and Ultra subscribers. Simon Willison notes that Neon makes the already confusing Gemini 4 pricing structure worse, and that the announcement omits Argon entirely; Latent Space reports Argon is now retired and that Neon is the fourth Gemini 4 model in about seven weeks.
Why it matters: Neon slots between Carbon and Hydrogen as the default choice for high-frequency agent and coding workloads, but the tier naming and relative costs shift again on December 1, so pin your model IDs and re-check which tier you are actually paying for.
Sources: blog.google / Simon Willison / Latent Space / Hacker News discussion
Anthropic prices IPO above range, shares begin trading today as ANT
Anthropic said it priced its initial public offering of Class A common stock, an offering of 980,000,000 shares, of which 956,351,354 are being sold by the company and 23,648,646 by selling stockholders. Shares are expected to begin trading on the Nasdaq Global Select Market on November 19, 2026 under the symbol ANT, with the offering expected to close on November 20. TechCrunch reports the price came in above the initial projected range and that the offering is the largest in history by proceeds raised, with Amazon holding nearly 15% of the company and Google almost 14%.
Why it matters: Anthropic's first day of trading becomes the public market's reference point for how much capital keeps flowing to frontier labs, and by extension how aggressively Claude's roadmap and capacity get funded.
Sources: Anthropic / TechCrunch AI / Hacker News discussion
OpenAI's GPT-6.2 Astra and Sol land cheaper, with sharply mixed early reviews
Latent Space reports OpenAI shipped GPT-6.2 Astra and Sol, available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users and through the API as gpt-6.2-astra and gpt-6.2-sol, both with 1M context and a 128K output limit. Artificial Analysis puts Astra at 62 on its Intelligence Index, behind Fable 5.2 at 65, while Sol matches September's Astra at a small fraction of the cost per task. Mercor's OpenWorldEngineer places Astra at 67.2% against Fable 5.2 at 54.6%. OpenAI rated Astra High for AI Self-Improvement, a first for one of its externally released models. Latent Space notes strongly negative influencer feedback, including Matt Shumer calling the models "essentially unusable" after weeks of testing.
Why it matters: Sol looks like the new default workhorse on intelligence per unit of spend, but the gap between benchmark results and hands-on reports means you should re-run your own evals before swapping it into production.
Sources: Latent Space
Microsoft releases seven MAI models at Ignite plus licensed weights for Azure
Microsoft announced seven new MAI models aimed at different price-performance points: MAI Code 2 and MAI Code 2 Flash for enterprise coding, MAI Cyber 2 Flash for defense work, MAI Image 2.7 and 2.7 Flash, MAI Realtime 1 for speech-to-speech, and Decision 1.1 for model routing. Microsoft says MAI Code 2 delivers 98% of GPT-6.1 Sol's performance and 96% of Claude Opus 5.5's across four coding benchmarks at a fraction of the price, and that developers preferred it over GPT-6 Sol 64% of the time in a blind frontend evaluation. The company also introduced MAI Licensed Weights, letting customers customize a frontier MAI model with proprietary data and run it inside their own Azure environment.
Why it matters: If your agents already run through GitHub Copilot or Microsoft Foundry, these are drop-in alternatives for high-volume routine work, and licensed weights are an option for teams that cannot send proprietary data to a hosted endpoint.
Sources: microsoft.ai / Hacker News discussion
Nvidia stops shipping licensed Hopper data center chips to China
The Verge reports that Nvidia has stopped sending advanced AI chips to China despite White House approval received in December last year. Interim CEO Colette Kress said on Tuesday's earnings call that guidance includes no licensed Hopper data center products and that the company is no longer shipping them to China; she did not explain why or say whether Nvidia would pursue other China business. The reversal follows Jensen Huang's departure as CEO last week. The Verge notes Reuters reported earlier this month that around 600,000 H200 chips were stuck in Hong Kong after Beijing withheld approval for large-scale imports.
Why it matters: Anyone counting on China-based GPU capacity or cross-border compute arbitrage should assume that supply is closed for now and plan capacity elsewhere.
Sources: The Verge AI
KPMG says AI agents asked sharper questions than its human audit team
KPMG UK ran more than 100 AI agents across the audit of an unnamed FTSE 250 company in a project codenamed Project Exponential, according to the Financial Times as reported by European Business Magazine. The agents collected information, analysed and documented data, drafted conclusions and suggested next steps while coordinating with each other. KPMG UK head of audit Cath Burnet said the technology went considerably deeper than the regular team in several instances and raised questions human auditors had not. The exercise was not a formally approved statutory audit, and an auditor had to intervene when one agent failed to correctly match regulatory requirements with company documentation. KPMG plans to expand to as many as 20 larger audits.
Why it matters: This is early field evidence that multi-agent systems can add value on judgment-heavy knowledge work, with the same caveats builders already know: data quality gates the whole thing, and a human still has to catch the regulatory mismatches.
Sources: europeanbusinessmagazine.com / Hacker News discussion
How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.
Sean Weldon builds apps and custom software at 22Atlas.