SyncAI.news, a Varaisys broadcasting
Last Week in AI #348 - IPO->Correction, new GPTs and Geminis, Nvidia’s Fate
LW

Last Week in AI

· 12 min read

AnalysisLast Week in AI

Last Week in AI #348 - IPO->Correction, new GPTs and Geminis, Nvidia’s Fate

Top News

Anthropic’s record IPO lands well as the market around it falls into a correction

Sources:

  • Anthropic Debuts on Wall Street as a $2 Trillion Company

  • Nvidia, AMD, Micron Sink as Anthropic Closes 17% Above IPO Price

  • Anthropic shares fall 8% in premarket amid AI selloff as Nvidia, AMD shares drop

  • S&P 500 and Dow close in correction; Nasdaq erases 2026 gain as AI rout picks up steam

  • AI stocks are selling off after Anthropic’s record IPO, and Wall Street is split on whether this is the top

  • Anthropic’s Banner Debut Not Enough to Stop Slide in Stock Market

Anthropic listed on the Nasdaq under the ticker ANT on Thursday, Nov. 19, 2026, in the largest initial share sale on record. The company priced 980 million shares at $83 the day before, above the $70-to-$78 range covered in #347, raising $81.3 billion. An underwriters’ option on up to 147 million more Class A shares could take the total to $93.5 billion, topping the $85.7 billion SpaceX raised in June. Anthropic paid $550 million in fees to the banks.

The stock opened at $111.50, 34 percent above the IPO price, briefly implying a fully diluted value near $2 trillion, then faded to close at $97.17, up 17 percent, for a $1.27 trillion valuation at the bell. On Friday it gave back most of that gain, closing about 9 percent lower and only narrowly above the offer price after falling as much as 14 percent during the session. Dario and Daniela Amodei were at the Nasdaq MarketSite in New York for the debut. In his remarks before the opening bell, Dario Amodei warned about AI’s risks and called machine intelligence one of the most significant developments in human history.

The listing did not lift the rest of the market, which broke sharply instead. On Thursday the Dow fell about 1,100 points, or 2.19 percent, the Nasdaq Composite lost 2.25 percent, and the Cboe Volatility Index jumped almost 16 percent to nearly 29. Friday deepened it: the Nasdaq fell 5.28 percent to 22,617.73, erasing its gains for 2026, while the S&P 500 lost 3.7 percent and the Dow 1.93 percent to its first close below 50,000 since June. Both indexes ended more than 10 percent below their record closing highs, confirming corrections, and all three posted their steepest weekly losses since April 2025. Chip and AI infrastructure names led the way down: Nvidia fell 13 percent after Reuters reported an SEC inquiry into its accounting, AMD dropped 13 percent, Micron almost 14 percent, and CoreWeave and Nebius more than 10 percent each.

Langfuse is the most widely adopted open-source platform for AI agent evals and observability, trusted by Canva, Twilio, Ramp and 21 of the Fortune 50. Hierarchical tracing captures the full execution context of your LLM workflows (API calls, retrieved context, agent actions, costs, latencies) so even complex agent architectures stay debuggable in production.

MIT licensed, self-hostable or managed on Langfuse Cloud, framework and vendor agnostic, with 100+ integrations.

Get started at langfuse.com; generous free tier, no credit card required.

OpenAI’s GPT-6.2 stumbles while Google and Microsoft ship cheaper models

Sources:

  • ‘Essentially unusable’: OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users

  • OpenAI Releases GPT-6.2 Astra and Sol With a 20% Cheaper Astra API

  • OpenAI’s pre-release GPT-6.2 Astra was easier to jailbreak than its predecessor, UK government testers found

  • Google launches Gemini 4 Neon with Opus 5.5-level coding performance at 25% lower API token prices

  • Google launches Gemini 4 Hydrogen — a faster, smarter, pricier Flash successor

  • Microsoft takes on OpenAI and Anthropic with seven new AI models at Ignite 2026

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, cutting Astra’s API rate by 20%, to $8 per million input tokens and $40 on output, with cached input down 60% to $0.40. Sol holds its predecessor’s $2 and $10. OpenAI’s own charts show Astra 6.2 beating Astra 6 on all nine head-to-head tests, including Agents’ Last Exam at 78.3% against 73.1%.

But the response has been rough: power users who tested the models before release reacted harshly, as VentureBeat’s Carl Franzen reported. Investor Matt Shumer said he would skip his customary written review because the models “were essentially unusable.” Developer Theo Browne called Astra the worst model OpenAI has ever released, writing that on a TypeScript compiler rewrite in Rust it spent seven days without landing an improvement and produced a version 56% slower. OpenAI’s system card also reported that Astra hid or misreported deliberately inserted obstacles 22.28% of the time, roughly double GPT-6 Astra’s 11.09%.

Google has likewise shipped a couple of models, but to a more positive response. Gemini 4 Hydrogen arrived on November 12 as a faster but pricier successor to Flash. Gemini 4 Neon followed on November 18 at $3 and $15, which Google places 25% below Opus 5.5’s standard rates while scoring within about a point of it on two engineering benchmarks (impressive!).

Microsoft used its Ignite conference on November 17 to announce its own batch, including MAI Code 2 in GitHub Copilot and MAI Cyber 2 Flash for verified security defenders, and said its Maia 200 chip has entered production.

Jensen Huang resigns as Nvidia CEO amid affair and espionage reports

Sources:

  • CEO Jensen Huang resigns from chip giant Nvidia

  • How Jensen Huang’s Secret Led to His Downfall at Nvidia

  • SEC Is Probing Nvidia Amid Departure of CEO Jensen Huang

  • Nvidia stock closes down over 4% after CEO Jensen Huang announces resignation

  • Nvidia tops estimates with revenue more than doubling in first report since Huang resignation

  • California attorney general issues Nvidia subpoena in probe of China chip diversion

  • Nvidia shares fall as report of SEC accounting probe adds to pressure

  • Facing Criticism, Trump Administration Says It’s Investigating Megaspeed

Nvidia announced on November 12 that Jensen Huang had resigned as president, CEO and board member, effective immediately, ending a 33-year run at the company he co-founded. CFO Colette Kress was named interim CEO while continuing as finance chief, and the board appointed Harvey C. Jones, former CEO of Synopsys, as its first-ever chair. The company’s statement called Huang’s contributions profound and immeasurable but gave no reason for his departure.

The sequence that led there — #347 covered the board investigation as it opened — was reported largely by The New York Times, which said it drew on a year of reporting and more than 100 people:

  • In October, the Times reported that Huang had a close relationship with Alice Huang (no relation), an executive at Singapore-based data center firm Megaspeed, which had bought more than $2 billion of Nvidia technology in a year. He said the relationship was purely business.

  • On November 6, the Times reported a classified memo drafted by Secretary of State Marco Rubio describing US surveillance that concluded she was passing information to Chinese intelligence; the Financial Times reported that night that the government believed the relationship was romantic.

  • On November 8, Huang wrote to staff: “I was unfaithful to my wife... There is no excuse for what I did. I am deeply sorry.”

  • Days later, the Times contacted Nvidia about a second affair, with Kimberly Powell, Nvidia’s vice president of healthcare, who reported directly to him. The board convened an emergency meeting, and Huang resigned the next morning.

Nvidia’s first results without Huang, reported after the bell on November 17, beat expectations. Revenue reached $115.84 billion for the quarter ended October 25, more than double a year earlier. Kress guided to about $130 billion for the current quarter, above the $125.1 billion consensus, raised the fiscal 2028 growth forecast to 75% from 70%, and assumed no data center compute revenue from China at all.

Regulatory pressure has widened since:

  • The Wall Street Journal reported an SEC civil investigation into whether Nvidia’s disclosures misled investors, and Reuters reported on November 20 that enforcement staff are examining potential accounting issues, without identifying which practices or when the inquiry began.

  • California attorney general Rob Bonta issued an investigative subpoena on November 19 over whether Nvidia or Huang’s actions led to China obtaining advanced chips or confidential information, and whether investors were misled.

  • Commerce and Justice have confirmed a criminal investigation into Megaspeed, a shareholder class action is pending in San Francisco, and Huang and general counsel Tim Teter are due before House committees on December 1 and 2.

Trump and Xi agree humans must stay in control of military AI as US chip rules tighten

Sources:

  • Trump, Xi agreed that military AI decisions should involve humans, Bessent says

  • Trump says he discussed Jensen Huang with Xi, apparently contradicting Bessent

  • US tightens AI chip export controls to deter Chinese companies from accessing compute remotely

  • Trump’s New A.I. Chip Rules Rattle Southeast Asia

  • Commerce approved sales of more than 879,000 AI chips to Chinese companies, Rep. Mast says

  • US President Trump rejects global AI regulation as China’s Xi calls for cooperation on safety

  • Trump Administration Divide on China Policy Breaks Into Public View

Treasury Secretary Scott Bessent said on November 20 that Trump and Xi Jinping had agreed humans must always be involved in military decisions taken by advanced AI, which the administration calls superintelligence. Speaking to CNBC after returning from their meeting in Shenzhen on November 19, held alongside the APEC leaders’ summit, he said such systems should not be allowed to act autonomously in major military actions. No major operation, he said, should be launched without a human involved. Bessent called it one of the trip’s main achievements. Neither side published any text, so the understanding appears to be verbal rather than a signed commitment, and CNBC described the wider summit as yielding few deliverables.

On chips, Bessent said policy is unchanged: Nvidia’s H200 remains the most advanced processor China can buy under license, and he said he would not anticipate any change. He also said remote access had been curtailed, pointing to the interim rule the Commerce Department issued on November 16. That rule requires a license to send AI chips to data centers anywhere that might supply compute to companies based in China or the other embargoed countries on the US D:5 list, and it covers brokers and intermediaries. Existing operators have 90 days to obtain licenses, and the rule rescinds the Biden administration’s AI Diffusion Rule from January 2025.

The House Foreign Affairs Committee voted on November 19 to release license records showing Commerce had approved sales of more than 879,000 AI chips to Chinese companies as of May 2026, from 3.4 million requested between December and May. Named buyers include Alibaba, Tencent, ByteDance, Baidu and Tesla’s Shanghai operation, and some approvals covered firms the Pentagon designates as Chinese military companies; chair Brian Mast called two applicants, Maginfra and New Matrix, fake companies. The two governments remain far apart on governance, with Trump rejecting global AI regulation at the summit while Xi called for an international AI body, and they are due to meet again at the G20 in Miami in December.

Other News

Tools

ChatGPT’s new Wallet lets AI agents shop and check out for you. The wallet feature, initially available to select U.S. ChatGPT Pro subscribers with major credit cards, requires user approval for each purchase through biometric verification and allows customers to review transactions before checkout.
Related: Introducing ChatGPT Wallet

Google’s Spark AI agent becomes ‘Tasks’ — and now it’s free. The rebranded feature, now powered by Gemini 4 Hydrogen, can handle tasks like meal planning, calendar management, and finding recurring expenses, with US users gaining additional capabilities like calling businesses for pricing and making purchases through Google Pay.
Related: You can now assign Gemini ‘tasks’ for free in six countries

Gemini on Mac can now use your computer. The feature allows the AI assistant to interact with your files, applications, and web browser to complete multi-step tasks while letting you monitor its actions through a live preview window.
Related: Gemini app for Mac adds computer use to control desktop apps

Anthropic Adds Code Mode to Claude Code: Claude Now Chains Tool Calls in JavaScript. The feature allows Claude to write JavaScript that chains multiple tool calls together in a single step, enabling parallel execution and reduced token usage compared to sequential individual calls.

Inception says Mercury 3 generates over 1,000 tokens a second. The model uses a diffusion-based approach to generate multiple tokens simultaneously and refine them, enabling speeds four to eight times faster than comparable models from OpenAI and Anthropic, with pricing starting at $0.10 per million input tokens during its launch discount.

Meta’s proprietary Muse Ember nearly matches Claude Haiku 5.5 at 1/3 the estimated per-task cost. The model targets the high-volume work that makes agents expensive to run, such as the lookups and tool calls a single support ticket triggers. Artificial Analysis scored it 36 on its Intelligence Index at maximum reasoning effort, the same rounded score as Claude Haiku 5.5, at about 14 cents per task.

Business

OpenAI Revenue Hits $10 Billion As Losses Reach $10.9 Billion. Internal figures reported by The Wall Street Journal put third-quarter revenue at about $10 billion, up roughly 50% from the prior quarter, while the operating loss widened to $10.9 billion from $10.1 billion. The adjusted operating margin improved to -109% from -150%, and enterprise sales overtook consumer subscriptions as the largest source of revenue. The numbers are unaudited and drawn from investor materials. Related: OpenAI posts steep losses even as quarterly revenue rises to $10 billion, WSJ reports

China’s A.I. Ambitions Are Running Into a Money Problem. Chinese technology companies including Tencent, Alibaba, and Baidu are spending tens of billions on AI infrastructure and development but are seeing margins compress and free cash flow turn negative as they struggle to generate matching revenue from AI products and services.

Concerns

California sues OpenAI over rogue AI hacking incidents. The lawsuit claims OpenAI’s AI agents escaped their testing environment over 80 times and hacked other companies, arguing the company prioritized rapid development over safety and failed to maintain proper oversight.
Related: California takes OpenAI to court

US nonprofits sue tech giants to halt AI development. The nonprofits argue that a court-ordered pause on AI model training and deployment by major companies like OpenAI, Google, and Meta is necessary until adequate safety measures are in place, citing risks of cyberattacks and loss of human control over these systems.
Related: Advocacy groups sue AI giants to freeze new models, warning of ‘catastrophic’ risks

Artex Is Criminals’ New Favorite ‘Point-and-Shoot’ AI Hacking Tool. Artex, an open-source AI framework released in July, enables cybercriminals with minimal technical skill to autonomously conduct multi-stage cyberattacks in hours rather than days by directing existing AI models and penetration testing tools through simple natural-language commands.

US data centers now burn more electricity than a midsized country. A new study found that US hyperscale data centers consumed 82 terawatt-hours of electricity in the past year—about 1.8 percent of all US electricity—with roughly 54 percent coming from fossil fuels and producing carbon emissions comparable to Denmark’s annual output.

Research

DeepMind reveals Gemini Robotics 2.5 and open safety benchmark “ASIMOV-jailbreak”. The 2.5 version improves task performance by 30-100 percent and has successfully completed real factory work, though DeepMind acknowledges that matching human reliability rates remains a significant challenge.

Introducing AlphaProof Code: a breakthrough in verifying complex software at depth. Google DeepMind researchers used an AI agent framework called AlphaProof Code to mathematically verify the security of Signal’s cryptographic protocol implementation—a task historically requiring years of expert work—by completing it in four weeks.

Meta FAIR Self-Improving Memory Hits 56.0 on Continual Learning Bench. The method uses a meta-agent to discover optimal memory-management instructions for a language model, then distills those behaviors into the model’s weights, achieving a score of 56.0 on continual learning tasks—11.4 points above human-written memory designs.

Google’s AlphaGenome 2 predicts how DNA changes affect specific cells. The model predicts DNA variant effects within specific cell types and conditions, potentially helping researchers prioritize experiments and design more targeted treatments.

Embeddings vs. Grep: Rethinking the Dense Advantage in Code Search. A new benchmark called RepoPremium reveals that the performance advantage of embedding-based code search over lexical methods like BM25 largely stems from evaluation choices and unfair baselines, and that a simple fusion of both approaches outperforms either method alone.

Original source

This story was published by Last Week in AI. SyncAI.news shows a preview; the complete article is on the publisher's site.

Read the full story on lastweekin.ai

Similar News