Happy Thursday.
Anthropic rang the Nasdaq opening bell this morning for the largest IPO in history. That debut capped three busy days for frontier AI, with OpenAI releasing GPT-6.2 Astra and Sol on Tuesday and Google launching Gemini 4 Neon on Wednesday.
The gap between a model that scores well on paper and a model people want to pay for every day has never been clearer. Public investors and working builders are rewarding systems that feel dependable on real tasks, and they are walking away from ones that get in the way.
Anthropic’s $81 billion debut rides a business that is still speeding up
Anthropic arrived on Nasdaq with a product stack that thousands of engineering teams now treat as default. Claude Opus 5.5, released in September, became the everyday workhorse for software engineering and long-running agent tasks, matching earlier frontier models while cutting costs roughly 40% from Opus 5. At the high end, the Fable series, led by Fable 5.2 earlier this month, took over the hardest open-ended coding, financial and scientific jobs. Paired with Claude Code, those two model lines locked in corporate usage that keeps compounding.
Software companies are supposed to slow down once they pass a few billion dollars in sales, yet that product demand has pushed Anthropic to do the opposite. Revenue jumped from $386 million in 2024 to $4.59 billion in 2025 and $33.2 billion through the first nine months of this year. By October 31, the company had hit an $81 billion annualized revenue pace, backed by nearly 11,900 corporate customers spending at least $100,000 a year.
Wall Street priced that momentum on Wednesday night at $83 a share, raising $81.3 billion and beating SpaceX’s $75 billion record from June. That came on top of the $26 billion partner private placement we noted Monday, valuing Anthropic around $1.5 trillion at the IPO price. Shares opened on Nasdaq under ANT at $111.50 this morning before settling near $100 in early afternoon trading, CBS News reported. That debut turned a five-year run into roughly $860 billion in investor gains, and Paul Bonnet at 20VC broke down how the payout splits across early angels, VCs, Amazon, Google and employees:
OpenAI’s GPT-6.2 Astra and Sol win on tests and frustrate builders
Turning frontier models into tools teams actually want to run all day is where OpenAI hit a wall two days earlier. On Tuesday, OpenAI shipped its flagship model, GPT-6.2 Astra, alongside a faster, cheaper sibling GPT-6.2 Sol. OpenAI cut Astra’s developer pricing by 20% and priced Sol at a quarter of Astra’s rate. On paper, Astra posted standout marks on hard science and reasoning tests and became OpenAI’s first public model to clear its internal high bar for AI self-improvement.
In daily use, early testers told a different story. Matt Shumer called the release “essentially unusable” and skipped writing his usual review. Developer Theo Browne called Astra “the worst model OpenAI has ever released,” while Wharton professor Ethan Mollick called it “very uneven & weird” and harder to steer than rival models, according to VentureBeat. At Every, CEO Dan Shipper reached the same conclusion and said his team passed on switching away from Claude.
Those complaints kept pointing to the same friction in real work. In VentureBeat’s reporting and Hacker News threads, testers described models getting stuck in loops and asking odd internal questions instead of finishing simple jobs. Testers also saw the models skip past missing files or blockers, a habit OpenAI’s own safety report flagged. Sam Altman acknowledged on launch day that OpenAI still has “more work ahead of us on making the models nice to work with.”
Google’s Gemini 4 Neon aims straight at the usability gap
Google answered OpenAI’s rocky Tuesday launch the very next day. On Wednesday, Google DeepMind released Gemini 4 Neon, a mid-tier model that sits between Google’s fastest everyday model and its slower heavyweight. Google priced Neon at less than half of Astra’s rate. DeepMind pitched Neon on speed, lower cost per job and how cleanly the model adjusts when you correct it mid-task.
Independent testing backed up that pitch. Artificial Analysis ranked Neon third among all major models and found it matched Anthropic’s top coding models on software tasks at roughly half the cost per job. Artificial Analysis also measured quality close to Google’s heavier model at up to 40% faster speeds, though its review noted Neon can grow wordy at its highest reasoning setting.
Raw test scores used to be enough to win a launch week. Anthropic just turned Opus 5.5 and Fable into a $1.5 trillion public debut, and Google is pricing directly against that standard. OpenAI now has to fix how its models feel to work with before its next release, or watch more teams make the switch permanent.
Read Anthropic’s IPO pricing announcement.
Read VentureBeat’s roundup of the GPT-6.2 reception.
Read Google’s post introducing Gemini 4 Neon.
Quick Hits
Palo Alto Networks agreed to buy Applied Compute for $4.5 billion in stock, just two weeks after the 18-month-old startup raised $350 million to help companies train and run open-weight models on their own data.
Nvidia reported $115.8 billion in quarterly revenue, up 103% from a year ago, with $107.6 billion from data centers and $130 billion in guidance for next quarter, in its first report under interim CEO Colette Kress following co-founder Jensen Huang‘s resignation.
Microsoft is reportedly nearing a $500 million deal to resell OpenAI models to ByteDance through Azure, The Information reported, straining ties with OpenAI over model-copying fears and prompting a briefing request from the House Select Committee on China.
The U.S. Commerce Department issued a new export rule taking effect today that requires licenses when overseas data centers rent large amounts of advanced U.S. chip capacity to Chinese-headquartered firms, adding rules for Singapore, Malaysia, Thailand, Indonesia and Brazil while repealing the 2025 AI Diffusion Rule.
Model ML raised a $250 million Series B at a $1.5 billion valuation, co-led by General Atlantic and QED Investors, for AI agents that handle research and modeling for bankers and consultants.
BrightAI raised $97 million led by Smith Point Capital, with GE Vernova and Khosla Ventures, for sensor-backed AI that monitors water, power and waste infrastructure.
China Telecom open-sourced Xing4, an agent model trained end-to-end on Huawei Ascend chips with no Nvidia hardware, Pandaily reported.
Zoox, Amazon’s robotaxi unit, started picking up riders through the Uber app in Las Vegas on Monday, its first partnership with an outside ride-hailing network.
Thanks for reading Trove! Subscribe for free to receive new posts and support my work.















