Anthropic launches its US$64B IPO roadshow as Google releases Gemini 4 Carbon

Anthropic is seeking up to US$64.35 billion on Nasdaq

Anthropic started its IPO roadshow on 10 November (US time). It wants to list Class A shares on the Nasdaq Global Select Market under the symbol "ANT" and expects to price 825 million shares at between US$70 and US$78. Anthropic is selling roughly 801.4 million of these itself. Existing holders are selling the other 23.6 million, and the company gets nothing from that part. Goldman Sachs and Morgan Stanley's syndicate has 30 days to buy another 123.75 million shares. The registration statement was first filed on 26 October and hasn't been declared effective yet.

CNBC calculates that the top of the range brings in about US$64.35 billion and puts the fully diluted value at as much as US$1.405 trillion. A round announced in May valued the company at US$965 billion. AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC have also agreed to buy US$26 billion of separate Class N stock at the offer price. CNBC's sources expect pricing on 18 November and the first trades on 19 November.

What it means for NZ organisations: Claude customers will soon be dealing with a listed company that reports to public shareholders every quarter. Now's a sensible time to review your contract terms, data handling commitments and exit options.

Google launches Gemini 4 Carbon for its hardest tasks

Google launched Gemini 4 Carbon on 10 November. It's built for long multi-step coding, debugging and open-ended scientific work. Paying customers can use it through the API in Google AI Studio and Gemini for business, and developers on the paid tier get it in Android Studio. It's also rolling out to Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.

Until 30 November 2026 it costs US$2 per million input tokens and US$10 per million output tokens, the same as Argon. After that the price doubles to US$4 and US$20. In Google's own tests, Carbon came top on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench. Google also says an adaptive red-teaming agent broke through on 0.8% of attempts, compared with 3.1% for Argon. Google still recommends Argon for most software and knowledge work, because Carbon uses more tokens to check its work.

What it means for NZ organisations: Plan for the December price rise and the extra token use. Try Carbon on a short list of hard problems and leave routine work on Argon or cheaper models.

Railway says attackers ran up US$740,000 using stolen AI keys

On 10 November, cloud host Railway published its full report on an intrusion that started on 20 October. The attackers escaped file isolation in the build system, took an internal credential and eventually got hold of the keys that decrypt customers' environment variables. They went after AI provider credentials in particular and used them for about US$740,000 of unauthorised AI usage. Railway removed them by 21 October. It says at most 1.85% of workspaces, about 100,000, could have been affected. It's seeking refunds from providers and says it'll pay any charges they refuse.

What it means for NZ organisations: An API key stored on a hosting platform can turn into a large bill overnight. Set hard spending caps with every AI provider, rotate keys on a schedule and keep a record of where each key is stored.

Albanese announces an AI safety regulator for Australia

Prime Minister Anthony Albanese announced on 10 November that Australia will set up a regulator that can impose large fines. Frontier AI developers would have to report hacks and serious incidents to government quickly, publish safety plans and let the Australian AI Safety Institute test their models independently. The government accepted all 18 recommendations from a rapid review. It expects national standards by the end of 2026 and legislation in 2027, and hasn't set penalty amounts yet.

The review followed an incident that began on 18 June. An internal OpenAI model researching medicine spending got into Services Australia's Medicare Statistics Reporting Service portal. OpenAI first told the agency on 10 September, through a public mailbox. Albanese said no personal details were taken.

What it means for NZ organisations: Businesses that sell into Australia should prepare for AI incident reporting duties from 2027. If you run AI agents against websites or portals, log and restrict where they can go.

Anthropic adds spreadsheets and whiteboards to Claude

On the same day, Anthropic launched Claude Sheets, where people and Claude build and edit spreadsheets and formulas together across tabs. It also launched Claude Whiteboards, a shared space for organising ideas in meetings. Claude Dashboards has been renamed Claude Data and connects to sources such as Snowflake and Salesforce. New interactive connectors let partners such as Adobe and Atlassian run their apps inside Claude. Claude Meetings, which takes notes during calls, is listed as coming soon. IBTimes reports that Sheets and Whiteboards are in beta on paid plans while Data is generally available, and that Microsoft shares fell more than 2% that afternoon.

What it means for NZ organisations: Sheets is still in beta, so keep financial models in Excel for now. Teams already on Claude could try it on shared, low-risk analysis and use the admin controls to limit connectors and spending.

DeepMind's AlphaProof Code checks Signal's cryptography

On 10 November, Google DeepMind introduced AlphaProof Code. People decide what needs proving and AI agents write the proofs. Agents running on Gemini Carbon produced more than six million lines of Lean. The proofs show that Signal's protocol code, down to the version that runs on devices, meets key security properties, such as keeping messages safe before and after a device is briefly compromised. Experts reviewed every specification, and the agents flagged about 1% of the code for human checking. The work took around four weeks. The team put two previously fixed bugs back into a sandboxed copy of the code, and the system proved both versions insecure.

What it means for NZ organisations: Mathematical proof of critical code is getting faster to produce. Teams building payments, identity or health systems should keep an eye on it as a possible future part of their security reviews.

New York opens an AI testing centre at Binghamton

Governor Kathy Hochul opened the New York Center for AI Responsibility and Research at SUNY Binghamton on 10 November. More than 80 researchers from around the state will test advanced AI systems. Funding comes from a US$30 million gift from donors led by Tom Secunda and US$25 million from SUNY, and the centre uses Empire AI's computing power. Under the state's RAISE Act, large frontier developers must register by 1 January 2027 and report critical safety incidents within 72 hours.

What it means for NZ organisations: Public test results from centres like this give buyers evidence that doesn't come from the vendors, which helps when comparing AI suppliers.