Anthropic's $64B IPO Roadshow, Gemini 4 Carbon Unleashes Agent

Dateline: SAN FRANCISCO

Anthropic has launched a roadshow for a proposed Nasdaq listing under ticker ANT, while Google opened access to Gemini 4 Carbon and its new multi agent coding workflow on the same day. The timing puts frontier AI, enterprise agents, and security costs on the same screen at once.

Why is Anthropic's IPO roadshow drawing so much attention?

Anthropic is seeking to sell shares at $70 to $78 each in a proposed Nasdaq listing under the ticker ANT. At the top of that range, the offering would raise about $64.4 billion, which would make it one of the largest IPOs ever attempted.

The filing covers roughly 825 million shares, including about 801.4 million new shares from Anthropic and around 23.6 million shares from existing investors. Seven tech giants have reportedly committed a combined $26 billion alongside the offering, a signal that major platforms see Claude as strategically important.

The roadshow matters because this would be the first time a pure frontier lab opens its books to public markets. Investors are likely to focus on inference margins, long term compute commitments to cloud partners, and the Long-Term Benefit Trust, which holds unusual sway over the board.

The deal is not final yet. The SEC registration is still not effective, so pricing and timing can still change before any listing.

What changed with Gemini 4 Carbon and Google's agent tools?

Google DeepMind has opened Gemini 4 Carbon to paid API developers and Google AI Ultra subscribers. The company is also previewing /swarm, a tool for coordinating multiple coding agents on the same job.

Carbon was described as Google's top tier model for harder, long horizon engineering work. It topped the Artificial Analysis Intelligence Index v5.0 with a score of 59, posted 88.2% on SEC-bench Pro for vulnerability discovery, and reportedly achieved a 0% success rate for indirect prompt injection attacks in Gray Swan testing.

The trade-off is clear. Carbon still trails some rivals on agentic coding indices and computer use benchmarks such as OSWorld, which is why Google's orchestration pitch matters: if one model is not best at every task, coordinate several models and combine their output.

Google also said pricing starts at an introductory $2 per million input tokens and $10 per million output tokens, with a later increase planned. In parallel, DeepMind unveiled AlphaProof Code, a multi model agentic system that generated more than six million lines of machine checked proofs and formally verified key security properties of Signal's encryption library in about four weeks, with humans reviewing roughly 1% of the output.

Who wins and who faces pressure from the agent economy?

The clearest winners are the platforms that can sell AI infrastructure, orchestration, and premium access. Anthropic could gain a public market valuation benchmark, while Google gains a stronger enterprise pitch by linking model quality with multi agent coordination.

Microsoft is also pressing its agent strategy. Satya Nadella has floated a future in which Microsoft's roughly 200,000 human workers operate alongside two million or even 20 million AI agents, and he says he personally runs around 100 agent sessions at once.

That vision lands next to hard budget and staffing realities. Microsoft is offering a voluntary exit package to around 60,000 employees, though it expects uptake below 5%, and Agent 365 is priced at $15 per user per month while capex is running at $45.2 billion a quarter.

Consumer health is also moving. Oura agreed to acquire Counsel Health, an AI assisted doctor service, for $625 million, and Bloomberg reported Counsel's fourth quarter revenue was up 153%. Wearable data tied to an AI medical front door gives Oura a more direct path into clinical workflows.

What does the latest security breach say about the downside?

It shows that agent economics can turn quickly into agent losses. Railway disclosed a breach in which attackers stole customer secrets that may have touched up to 100,000 workspaces, and the stolen credentials were reportedly used to rack up about $740,000 in unauthorized AI charges.

That is the operational risk behind the enthusiasm. If model access keys are compromised, billing can spike fast, and the damage can run far beyond a single account.

The response is already forming. The Linux Foundation launched the Agent Name Service, an open project to register and verify AI agent identities using DNS records and certificates, with 11 early supporters.

The broader market message is direct: the next phase of AI is less about isolated demos and more about coordination, security, and payment rails. Anthropic's roadshow will test how much public investors will pay for frontier risk, while Google, Microsoft, Oura, and security teams all build around the same central fact, agents are becoming expensive, useful, and vulnerable at the same time.