Skip to content
AI Workshop

Anthropic Closes 17% Up After the Biggest IPO Ever, GPT-6.2 Matches the Top-Tier Price, Nvidia Loses Its CEO — and Microsoft Gives AI Agents an Identity

Anthropic is now a listed company. Its shares were priced above the range, raised more money than any IPO before it and closed the first day 17% higher. In the same week OpenAI matched Anthropic's top-tier price, Microsoft gave AI agents their own identities, and Nvidia reported record revenue a few days after losing its CEO. For companies that use AI, the useful thread is simple: top models now cost about the same everywhere, so the decision is about fit, governance and fallback, not price.

OpenAI released GPT-6.2 Astra and Sol, with Astra 20% cheaper and priced like Claude Fable 5.2. Google's Gemini 4 Hydrogen brought the first Gemini 4 model to developers at a lower price point. At Microsoft Ignite the theme was managing agents like employees: an identity, a control plane, and a cloud PC of their own. Anthropic launched Claude Meetings and started watermarking text from some models, two weeks before the AI Act's marking deadline. At home, Hypefy ranked sixth on Deloitte's Central Europe Fast 50, and Zagreb's startup week is coming up.

The Releases at a Glance

GPT-6.2 Astra · Sol
OpenAI · 17 November

Astra for hard work such as software development, Sol for everyday tasks and multi-step agents. Reported API prices: Astra $8 / $40, Sol $2 / $10 per million input / output tokens.

Gemini 4 Hydrogen
Google · 12 November

The fast, low-cost Gemini 4 model, in the API, AI Studio, the Gemini app and Search. $1.50 / $7.50 per million tokens; Google says it is 2.5 times faster than Gemini 3.8 Flash.

Claude Meetings · self-hosted Managed Agents
Anthropic · 12 November

Claude joins Google Meet calls as a visible participant and turns the discussion into notes, tasks and documents (beta, Team and Enterprise). Managed Agents can now run on your own infrastructure.

Agent 365 · Entra Agent ID
Microsoft Ignite · 17 November

A control plane for every agent in the company, agent identities in Entra (generally available), Copilot Workspaces where people and agents share projects, and Windows 365 for Agents.

Anthropic's IPO: Priced High, Closed Higher

Last week Anthropic set a range of $70 to $78 per share. On 18 November it priced the offering at $83, above the range, raising about $81.3 billion, the largest IPO on record according to Yahoo Finance and BNN Bloomberg. Trading under the ticker ANT began on Nasdaq on 19 November. The first trade was at $111.50, the stock touched $115.10 and then drifted down to close at $97.17, about 17% above the offering price. Yahoo Finance put the valuation at about $1.27 trillion at the close; other outlets count it differently.

For customers, the important part comes later: a listed Anthropic has to publish quarterly figures on revenue, costs and its compute commitments. That makes it easier to judge how stable a supplier is. It changes nothing about the models or contracts you have today, and nothing here is investment advice.

The Top Tier Now Has One Price

Two weeks ago the most capable models from the big labs had different prices. Now they don't. According to MLQ.ai and other coverage, GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than its predecessor and the same as the reported price of Claude Fable 5.2. Sol stays at $2 / $10. OpenAI's model IDs are gpt-6.2-astra and gpt-6.2-sol, and both appear in ChatGPT for paid plans and in Codex. One source reports a context window of about one million tokens; treat that as unconfirmed until you see it in your own console.

One level down, Gemini 4 Hydrogen costs $1.50 / $7.50, more than the Flash model it replaces, and Google claims it is 2.5 times faster and much better at coding (57% on Terminal-Bench 4.1 against 18.2% for Gemini 3.8 Flash, Google's own figures). The heavier Gemini 4 models, Carbon and Argon, are not generally available yet.

What it means for a normal company: every major vendor now sells a top model at about $8 / $40 and an everyday model at a fraction of that. Price no longer decides which vendor you choose. What decides is which model passes your tasks, where the data is processed, and whether you can switch if a supplier changes terms. Route most work to the cheaper tier and send only the tasks that fail there to the top tier.

Microsoft Ignite: Agents Get an Identity

Satya Nadella called Copilot "the new operating system for work", but the more practical news was about control. Entra Agent ID is generally available, so each agent can have its own identity with permissions and an audit trail, like an employee account. Agent 365 is the control plane that lists, governs and restricts agents across the company; it costs $15 per user per month on its own, or comes in the $99 Microsoft 365 E7 bundle, according to licensing analysts. Copilot Workspaces lets people and agents work on shared projects, Windows 365 for Agents gives an agent its own cloud PC to operate old business software, and Foundry IQ adds knowledge graphs and permission-aware SharePoint retrieval for agents. On Azure, Azure Catalyst combines Azure Migrate with GitHub Copilot agents to modernise applications.

"Generally available" does not mean every tenant has it today, and autonomous agents are still in Microsoft's Frontier early-access programme. But the direction is clear: in a Microsoft shop, an agent will be a managed identity, not a script running under someone's personal account.

Nvidia: Record Quarter, No CEO

On 12 November Nvidia announced that Jensen Huang had resigned as president, CEO and board member, effective immediately, after 33 years. The resignation followed an independent investigation that the board had opened a few days earlier. CFO Colette Kress is interim CEO, and former Synopsys chief Harvey C. Jones became the company's first board chair. The board is searching for a permanent successor.

Five days later the company reported its fiscal third quarter: revenue of about $115.8 billion, up 103%, with $107.6 billion from data centres. Guidance for the next quarter is about $130 billion, and it assumes no data-centre compute revenue from China. Demand for AI hardware is not slowing, whoever runs the company.

Watermarks, Privacy and Safety Checks

AI Act

Anthropic starts watermarking text

Since 16 November Anthropic's developer platform watermarks text from selected models, according to its release notes; non-default sampling settings can switch it off. The timing matters in Europe: from 2 December, generative AI systems already on the market must mark their output in a machine-readable way under Article 50(2) of the AI Act. If you build products on these APIs, check which of your models and settings carry the mark, and do not strip it in post-processing.

Privacy

Gemini separates history from training

Google is adding separate privacy settings to Gemini, so users can keep their chat history without it being used to train models, and AI Mode in Search gets temporary chats that are not saved. For companies whose staff use consumer accounts, this removes one excuse but not the main problem: business data belongs in a business plan with a contract.

Safety

OpenAI invites outside evaluators in

OpenAI announced that staff from four AI safety and research organisations will work inside the company as "embedded" evaluators. NPR reports researchers' doubts that this is enough, and OpenAI itself says there is no standard yet for this kind of evaluation.

Tutorial of the Week: Give Every Agent a Name and an Owner

Ignite made agent identity a product feature, and the AI Act will soon ask who is responsible for which system. You can start this week, whatever platform you use.

  1. List every agent and automation that calls an AI model: chatbots, Copilot agents, n8n or Zapier flows, scripts with an API key. Ask each team; the shadow ones are the ones that matter.
  2. Assign one human owner per agent, by name. If nobody wants to own it, switch it off.
  3. Move each agent off personal accounts and shared keys onto its own identity: an Entra Agent ID, a service account or a separate API key per agent.
  4. Grant the minimum permissions: read-only by default, write access only to the systems it really needs, and no access to payroll or HR data unless that is its job.
  5. Log what it does: prompts, tool calls and actions, kept long enough to investigate a mistake.
  6. Write down the model, the vendor and the fallback for each agent. With top-tier prices now equal, a second vendor costs little to keep ready.
  7. Review the list every quarter. Agents multiply quickly; the list is your AI Act inventory in its first form.

Croatia Corner: Hypefy, Hotels and Startup Week

Growth

Hypefy is sixth on Deloitte's Central Europe Fast 50

Zagreb-based Hypefy, which automates influencer marketing from selection to payment, grew revenue by 3,112% between 2022 and 2025 and ranked sixth on Deloitte's Technology Fast 50 for Central Europe, Netokracija reports. It now operates in 45 countries, and in October its revenue from Germany, Austria and Switzerland passed its revenue from the home region for the first time. The company set up an internal AI Enablement team to redesign work across departments. Nine Croatian companies made the list, including Orqa, Farseer, DICE, Udonis and Monri Payments.

Tourism

Hotel websites are not ready for AI agents

In a Netokracija column, Domagoj Razum looked at 242 hotel brands and found their websites are built for people, not for the AI agents that increasingly search for the best offer on a guest's behalf. For a country that lives from tourism, that is a practical to-do before next season: clear prices, availability and policies that a machine can read.

Education

A free AI handbook for teachers

The CroAI association published a free digital handbook on using AI thoughtfully in Croatian schools, with 28 chapters and 88 ready-made prompts and examples from Croatian classrooms, Bug.hr reports. It fits the national AI plan, whose public consultation closed on 15 November and which aims to train 10,000 teachers.

Events

NaPOD on 25 November, Zagreb Connect on 26 November

Zagreb Connect at ZICER hosts the final of the city's acceleration programmes, for which Zagreb has set aside €300,000, and an open pitch for startups from all over Croatia. NaPOD takes place in Zagreb the day before.

The Week in Numbers

$83
Anthropic's IPO price, above the $70–78 range
+17%
ANT at the first-day close ($97.17)
$115.8B
Nvidia's quarterly revenue, up 103%
3,112%
Hypefy's revenue growth, 2022–2025

What to Watch

  • 25 November: NaPOD in Zagreb.
  • 26 November: Zagreb Connect at ZICER.
  • 2 December: AI Act deadline for machine-readable marking of output from generative AI systems already on the market.
  • Coming weeks: Nvidia's search for a permanent CEO, and a wider release of Gemini 4 Carbon.
  • Early 2027: Anthropic's first quarterly report as a listed company.

Main sources: Yahoo Finance, BNN Bloomberg, MLQ.ai, ComputingForGeeks, VentureBeat, TechCrunch, Claude Platform release notes, Virtualization Review, Constellation Research, Next.ink, NVIDIA 8-K, The Next Web, Let's Data Science, NPR, Netokracija on Hypefy, Netokracija on hotel websites, Bug.hr on the CroAI handbook, Forbes Hrvatska.

Same price everywhere: now choose on fit

We test models on your real tasks, set up a second vendor as a ready fallback, and give every agent an identity, an owner and a log, so you are ready for the AI Act inventory. Claude Certified Architect, based in Zagreb.

Talk to us

Boris Agatić

Founder of AI Workshop and Claude Certified Architect. Over 10 years of building software, guiding companies from the first strategy call to production AI agents.

LinkedIn (opens in a new tab)