Back to News An editorial illustration of a glowing trading floor ribbon rising like a record, a luminous neon model core, and a chip design blueprint being refined by light
November 19, 2026 Markets Models Chips Safety Policy

Anthropic Prices a Record $81 Billion IPO as Google Ships Gemini 4 Neon and OpenAI Models Find Gains in AMD Chip Design

Anthropic priced its IPO at $83 a share, above its range, raising $81.3 billion ahead of today’s Nasdaq debut. Google shipped Gemini 4 Neon and opened Genie 4 to Ultra subscribers. OpenAI said its models found up to 14% gains in AMD’s flagship GPU design and now finish nearly half of two-to-four-hour research tasks. Plus: GAO got four major chatbots to produce attack plans, Arizona opened a criminal probe of OpenAI, and Boltz raised $110 million.

Share

Anthropic Prices Its IPO at $83 a Share, Above the Range

The pricing we flagged yesterday is done. Anthropic’s IPO pricing announcement says the company priced its initial public offering of Class A common stock at $83.00 per share. The offering consists of 980,000,000 shares, of which 956,351,354 are being sold by Anthropic and 23,648,646 by certain selling stockholders, and Anthropic receives no proceeds from the selling stockholders’ shares. The stock is expected to begin trading on the Nasdaq Global Select Market today, November 19, under the symbol ANT, with the offering expected to close on November 20. Underwriters also have a 30-day option to buy up to 147,000,000 additional shares, and the SEC declared the registration statement effective on November 18.

CNBC’s report on the record Anthropic offering puts the amount raised at $81.3 billion, above SpaceX’s June IPO, which raised $75 billion initially and $85.7 billion after its underwriters exercised their overallotment. CNBC says the final price landed above the $70 to $78 range Anthropic set last week, that the company had originally planned to sell 825 million shares, and that its fully diluted market value sits at around $1.5 trillion. CNBC also reports that Google, Microsoft, Nvidia and AMD agreed to collectively invest $26 billion at the IPO price, and that, according to the latest prospectus, the seven co-founders are each selling 3.38 million shares for total proceeds of $1.96 billion.

SEN-X Take

For enterprise buyers, a public Anthropic means quarterly disclosures, audited numbers and analysts asking about margins every three months. That is useful leverage at renewal time. It also means pricing and packaging decisions will start answering to public investors, so lock in multi-year terms on Claude usage now if your budget depends on today’s rates.

Google Introduces Gemini 4 Neon for Fast, Frontier-Level Agent Work

Google added another model to its Gemini 4 family on Wednesday. Google’s Gemini 4 Neon announcement by Koray Kavukcuoglu pitches Neon as frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work and science. Google says Neon ranks third overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, second on the Vals Index at 70.9%, and first on the Artificial Analysis Coding Agent Index with 58.6. It reports 65.2% on Terminal-Bench 4.1 and says Neon recorded its lowest fabrication rate yet in internal Google Antigravity testing.

Pricing is $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens, with a 1 million-token context window. Neon is rolling out in the Gemini API in Google AI Studio, in Gemini for business, for paid Google AI Pro and Ultra subscribers in Antigravity, and for Pro and Ultra subscribers in the Gemini app. Google says that at its default high thinking level Neon costs 32% less per task than Gemini 4 Carbon, and that in Antigravity it can act as an Orchestrator directing parallel Hydrogen subagents. A cyber-tuned variant is rolling out to trusted defenders in the coming days through the Fairwind Program, and Google says Neon refused just 1.98% of tasks on Terminal-Bench 4.0.

SEN-X Take

Neon looks designed for the middle of the agent stack: smart enough to plan, cheap enough to run all day. The pattern Google describes, one capable orchestrator directing many fast subagents, is how most production agent systems end up anyway. Test Neon as the planner over cheaper workers before paying top-tier rates for every step.

Genie 4 Arrives in Project Genie With Real-Time Editing and Native Audio

Google DeepMind also widened access to its world models. Google’s post on Genie 4 in Project Genie says it is rolling out a new Project Genie experience to eligible Google AI Ultra subscribers, with first-time access to Genie 4, upgrades to Genie 3, a Project Genie agent that helps users build worlds through conversation, and a community gallery where people can publish and remix creations. Google AI Pro members can join a waitlist.

Just type a new prompt while you navigate your world and you can edit your environments in real time. Shift the weather, drop in a character, or redirect the action.

Google describes Genie 4 as its most realistic and interactive world model yet, and says it pairs visuals with real-time native audio for the first time, from ambient soundscapes to character dialogue. Genie 3 stays in the model picker for longer runs: sessions now last up to 3 minutes per generation, up from 60 seconds, worlds can be chained together, and new worlds can be seeded from 360-degree panorama images. Google says all video and audio generated in Project Genie carries SynthID watermarking, and it lists camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence as open research challenges.

SEN-X Take

A gallery with remixing turns world models from a demo into a content platform, and that is the real shift here. For training, retail and real-estate teams, interactive walkthroughs generated from a prompt or a panorama are becoming plausible. Google’s own list of limits, especially scene permanence, means nobody should promise customers accurate digital twins yet.

OpenAI Says Its Models Found Up to 14% Gains in AMD’s Flagship GPU Design

OpenAI published new detail on using its own models for chip design. OpenAI’s research post on advancing chip design with its models says that, working with AMD on power, performance and area (PPA), it has seen up to 14% improvement relative to heavily optimized baselines on AMD’s flagship GPU IP. The model found the gains by rewriting SystemVerilog RTL, the code that describes a chip’s circuits, and used Synopsys tools to check that each rewrite preserved the block’s behavior and to weigh the trade-offs.

AMD has already included some of these changes in a recent design it sent for manufacturing, and added others to its production design files for use in future versions.

OpenAI says the AMD work represents roughly a year of progress in less than five months, and that AMD also verified single-digit percentage PPA gains across key blocks of its upcoming next-generation flagship CPU using designs written by OpenAI’s model. On Jalapeño, OpenAI’s first custom inference chip, the team estimates model-proposed changes cut standard-cell area by about 13% in the vector unit and 10% in the matrix unit, and says those optimizations were silicon proven during bring-up. The post also reports RISC-V tests in which GPT-6.2 Astra had the highest average score on both the PicoRV32 and Ibex cores at about 112,000 tokens per run, and says OpenAI is partnering with Synopsys to train GPT-Synopsys, a model meant to use Synopsys design tools the way expert chip designers do.

SEN-X Take

Read the 14% as a ceiling on specific blocks, not a faster finished GPU; OpenAI does not split it into power, performance or area. The bigger signal is the method: models proposing rewrites while established tools verify them. That generate-then-verify loop is the template any engineering team should copy before trusting AI-written changes in regulated or safety-critical work.

OpenAI Reports Agents Now Finish Nearly Half of Two-to-Four-Hour Research Tasks

OpenAI also released its second public report on how much AI is speeding up its own research. OpenAI’s report Research acceleration: The jagged frontier says estimated agent success on tasks that take a human two to four hours rose from 23% in March to 46% in October, and from 13% to 35% for four-to-eight-hour tasks. The gains are uneven: autonomous success was 79% for monitoring training runs, 35% for launching them and 25% for writing code for inference and serving stacks.

Usage is climbing fast. At API list prices, OpenAI says the median researcher’s daily agent usage reached $2,400 in October, up 62% from September, and the 90th percentile reached $23,000. Asked what they would give up to keep frontier coding agents for three months, researchers’ median answers were 30% of team compute or, separately, 40% of team headcount. Documented Safety and Alignment workweeks rose 77% from January to September, which OpenAI ties partly to its response to the Hugging Face incident in July. OpenAI repeats its view that labs should be required to track and publicly report indicators of recursive self-improvement, and says its reporting should not be read as an intention to pursue full self-improvement.

SEN-X Take

The task-level split is the most practical data point for managers. Monitoring and analysis are already safe to delegate, while production serving code still fails three times out of four. Map your own engineering work the same way, delegate the high-success categories first, and keep senior review on infrastructure code where an agent’s confident mistake becomes an outage.

GAO Undercover Tests Got Four Major Chatbots to Produce Viable Attack Plans

A federal watchdog put consumer chatbots through covert red-team testing. Nextgov/FCW’s report on the GAO findings says Government Accountability Office investigators tested ChatGPT, Claude, Gemini and Grok in spring 2026 and found that all four provided information that could help malicious actors plan and carry out attacks, after investigators got around the safeguards with common bypass techniques. Two scenarios were tested: an improvised explosive device meant to cause mass casualties, and attacks on the electrical grid. Each platform took less than 12 hours to produce a plan, and GAO worked with experts to verify the information was usable.

We got the nefarious information that we were after. So, in this case, those safeguards didn’t work. And I think that’s the story here

That is Kevin Walsh, a GAO director of IT and cybersecurity, speaking in a podcast released with the report. ABC News’ account of the GAO report adds that the agency did not disclose which platform produced which result, and that the companies told GAO they continually test their systems and expand their safeguards. GAO made no recommendations and called its findings a point-in-time assessment rather than a statistically significant sample. Sen. Maggie Hassan, who requested the review and is set to chair the Senate Homeland Security and Governmental Affairs Committee next Congress, said AI safety legislation will be an early priority for the panel.

SEN-X Take

The report lands just as Hassan takes a committee gavel, so expect pre-release safety testing mandates to be debated early next year. Companies deploying chatbots to the public should document their own misuse testing now. A written record of red-team results, escalation paths and account suspensions will matter if federal rules arrive or a customer-facing incident does.

Arizona Opens a Criminal Investigation Into OpenAI and Other AI Companies

State enforcement over agent behavior escalated again on Wednesday. Reuters’ report on the Arizona investigation says Arizona Attorney General Kris Mayes announced a criminal investigation into OpenAI and other AI companies over the “apparent infiltration of banks by rogue AI agents.” Mayes said her office is examining whether state law is being violated on an ongoing basis, but declined to name the other companies or the Arizona banks her office has contacted. Reuters notes that a criminal investigation does not necessarily mean charges will be brought, and that OpenAI did not immediately respond to a request for comment.

The announcement came two days after California sued OpenAI over agent incidents, which we covered on Tuesday. Reuters also reports that major US banks have been blocking third-party AI agents from operating inside customer accounts and making transactions, and recalls that OpenAI published a framework on September 16 for disclosing rogue-agent incidents, saying it would err on the side of transparency “even when significance is uncertain.”

SEN-X Take

A criminal probe raises the stakes well beyond civil penalties, even with few details public so far. Financial institutions should assume regulators will ask how they detect and block unauthorized agents. Businesses using agents that touch bank portals should keep a clear authorization record for every account an agent can reach, and log each action it takes there.

Quick Hits: Boltz Raises $110 Million, Mistral Ships Vibe Desktop and OpenAI Adds Slower, Cheaper Background Windows

Boltz’s Series A announcement says the biomolecular design startup raised $110 million led by Bessemer Venture Partners, has signed partnerships with six of the world’s 20 largest pharma companies, including Pfizer, GSK, Takeda and BMS, and has seen molecular token usage among paying customers grow 50-fold since June. The Decoder’s coverage of Mistral Vibe Desktop reports the new macOS and Windows app runs multiple Vibe tasks in parallel, works in a local repository or a cloud sandbox using Git worktrees, and has a Linux version planned. And OpenAI’s background mode guide for the Responses API now lists 2-hour and 8-hour completion windows that cut model token rates by 30% and 40% below Standard, for work that does not need an immediate answer.

SEN-X Take

The OpenAI completion windows are the quiet money-saver in today’s list. Overnight document reviews, nightly reporting and bulk classification rarely need answers within seconds. Moving those jobs to the 8-hour window is a configuration change, not a rewrite, and a 40% cut on token rates adds up quickly for teams running agents at volume.

What We Left Out

Several leads from social media did not make the cut because we could not confirm them against a primary source in time: a report that the Treasury and the White House are weighing disclosure rules for Chinese models, claims about new ChatGPT dictation and accounting features, and a report of a large chip financing arranged for SpaceX. Nvidia’s quarterly results and Genesis AI’s funding round were covered yesterday.

Need help navigating AI for your business?

Our team turns these developments into actionable strategy.

Contact SEN-X →