A big day for the business of AI, plus one wild research result. Each explained from scratch, no assumed background, with links to go deeper.
Illustration.
1Anthropic updates its IPO paperwork, and the numbers are striking
Anthropic, the company behind Claude, is preparing to sell shares to the public for the first time (an IPO). On Monday it updated its official filing with fresh financials: about $33 billion in revenue over the first nine months of 2026, up roughly twelve-fold from a year earlier, and $17 billion in the third quarter alone. The eye-catching figure is on the other side of the ledger: about $895 billion in future spending commitments, almost all for computing power - promises to pay for data centers and chips years into the future.
The filing also includes an unusual safety disclosure: Anthropic says its internal estimates suggest future models may "imminently" cross a threshold where AI starts meaningfully speeding up AI research itself, and that it has sometimes noticed such acceleration only after it happened. It says it is weighing new rules to keep humans in control. This is the company's own assessment of its future models, not an established finding about progress that has already happened.
Why it matters: this is the clearest look yet at the economics of a frontier AI company - enormous real revenue and even larger future costs, and unusual candor about where it thinks progress is heading.
2Salesforce agrees to buy an AI call-center company
Salesforce, whose software runs customer service for thousands of companies, agreed to acquire Cresta, a startup whose AI "voice agents" can handle entire customer phone calls - naturally, with interruptions, at scale - for brands like Marriott and Comcast. The price wasn't officially disclosed; The Information reports about $2.7 billion. The deal still needs to close.
Why it matters: phone support is where AI meets ordinary people most directly. It's a sign the biggest customer-service platform sees voice agents as ready for prime time, though the deal still needs to close and rollout plans aren't public. For the industry, it's another sign the action is moving from chat windows to real conversations.
An MIT-affiliated team posted a strange result: they handed a frontier AI model the raw internal settings (called "weights") of a much smaller AI model and asked it to figure out why the small model was bad at math. The frontier model identified a flaw and rewrote about 0.016% of the weights directly - no retraining run, no new data - lifting the small model's math score from 7.9% to 48.1% on a standard test.
Why it matters: today, improving an AI model means expensive retraining. If an AI can simply read another model's wiring and patch specific weaknesses, fixing problems could become fast and surgical. Important caveats: this is a blog post, not peer-reviewed science, and it worked on a small research model, not a frontier one. Treat it as an intriguing early result, not a capability you can use.
Checked against original announcements and named press reports on the morning of November 3, 2026. Anthropic's IPO and the Salesforce-Cresta deal are not final, and the MIT result is unverified early research.
