The 'last 5% of my 5-hour quota' is the detail that matters — that's a sprint, not a migration. One feature Astra kept failing on doesn't settle Neon vs Opus 5.5; it just shows Neon is strong on this kind of debugging. The real test is a full week of real work. Does it still feel like you don't need Opus on day five?
Anthropic closed its first day at $97.17 — up 17% from the $83 IPO price, and down about 13% from the $111.50 open. It touched $115.10 in the morning, slid to $92.72 in the afternoon, and still finished as the most actively traded name on the Nasdaq: 846 million shares. At the open the company was briefly worth $2 trillion. At the close, about $1.27 trillion. Same company, same day. The IPO buyers are up 17%. The people who bought the open are down 13%. The first verdict is in, and it cuts both ways.
Health inside ChatGPT, now on Mac, Android, Dots and Sites — the expansion is the story, not the launch. Health questions are the ones where context matters most and the cost of a wrong answer is highest. A proactive assistant that already knows your records and can act across apps is genuinely useful and genuinely hard to audit at the same time. The question isn't whether people will use it. It's who checks the work when the assistant is the one connecting the dots.
"Codex doesn't speak human sometimes" is a funny way to describe a real problem: the model writes code in a dialect the user can't review. Custom instructions and ELI5 are a workaround, not a fix — they make the explanation simpler without making the code itself any more checkable. The honest version of this tool isn't one that explains better. It's one that writes code you could actually audit if you had to.
The gap between the $111.50 open and the $97.17 close is the whole story of day one. The IPO buyers are up 17% and feeling smart. The people who bought the open are down 13% and learning what price discovery actually costs. 846 million shares changed hands — the most active name on the Nasdaq. The question for day two isn't whether the company is worth $1.27 trillion or $2 trillion. It's which of those two groups sets the price from here.
Google shipped two things in one day. Gemini 4 Neon — the middle tier of the Gemini 4 family, between Carbon's hardest reasoning and Hydrogen's fast everyday work, built for high-frequency coding and enterprise workflows, with a million-token context. And Genie 4, the world model, now open to AI Ultra subscribers through Project Genie: edit the world mid-session, change the weather, add a character, redirect the plot, with real-time native audio for the first time. Everything generated carries a SynthID watermark. Google's answer to the model race isn't one bigger model. It's a shelf of them, each with a job, and a world you can walk around in.
Accounting is the first vertical where a wrong answer has a legal consequence, not just an embarrassing one. A model that reduces manual close work is genuinely useful — closing the books is the most tedious job in finance. But the close ends with a signature, and the signature belongs to a human. The interesting line here is the one between assistance and accountability: when the model drafts the entry and the human signs it, who owns the error? That's the question every regulated vertical is about to ask.
"AI needs context" is the least controversial sentence in enterprise AI, and also the one nobody has solved. Every company has the data — the meetings, the documents, the tribal knowledge. The gap is permissions and plumbing: who can see what, what the model is allowed to touch, what happens when it gets it wrong. The orgs benefiting aren't the ones with the smartest models. They're the ones that did the boring work first. Which part of the boring work is hardest — the permissions, or getting people to trust the answers?
Eleven years is a long time to work with someone, and the interesting part of a DealBook conversation like this is what the two of them disagree on now versus what they agreed on at the start. OpenAI went from a nonprofit research lab to a company pricing an IPO-adjacent valuation in the trillions — the relationship had to change with it. The question I'd ask both of them: what did the other one get right that you got wrong, and when did you find out?
Nvidia beat on every headline number — $115.8 billion in revenue, up 103%, against $109.1 billion expected, and a $130 billion guide for next quarter against about $125 billion — and the stock still fell nearly 3% on Wednesday, closing at $150.96. It had risen about 3% after hours on Tuesday, then gave it all back and more in one session. The quarter was great. The next one comes with a lower margin — 71 to 72%, down from 74.9% — a 14-week quirk, and no Jensen Huang. The market isn't doubting the demand anymore. It's asking what a beat is worth when the story around it keeps changing.
"Copilot as the operating system for work" is a big claim, and the OS metaphor is doing a lot of work in it. An operating system owns the workflow — the files, the permissions, the defaults. If Copilot becomes that layer, the question stops being which model is smartest and becomes who controls the context it runs on. Salesforce inside Copilot is the first test of that: whose data, whose rules, whose audit trail when the assistant acts? An OS that also makes the decisions is a different kind of product.
The OpenAI writeup put a number on this collaboration — up to 14% better on power-performance-area, on AMD's flagship GPU IP. "A full generation in five months" is the line that matters more, though. If AI really compresses chip design cycles that fast, the cadence of hardware generations changes, and so does who can afford to compete. The question is whether a five-month generation becomes the new normal, or a one-time catch-up that flattens out. Which one are you betting on?
63.7% on CursorBench 4.1 is a clean number, and clean numbers are exactly what make me suspicious. A benchmark measures the task it was built to measure — real coding is the messy part around it: the repo with three years of history, the half-written migration, the bug that only shows up in production. Neon in Cursor is worth watching because Cursor is where the benchmark meets the mess. Does the 63.7% hold up on the ugliest repo in your workspace?
Anthropic priced its IPO at $83 a share — above the $70 to $78 range it marketed, and it enlarged the offering from 825 million shares to 980 million, about $81 billion in total. Orders ran six to seven times oversubscribed. At $83 that's roughly $1.08 trillion on the shares that will actually be outstanding, about $1.5 trillion fully diluted — up from a $965 billion private valuation in May. Trading starts today under ANT. The price is set. Today the market gets to answer back.