# The Edge: Claude joins your Google Meet and starts the plan before you hang up. Plus a free Gemini 4

> For teams on Claude Team or Enterprise: what it adds over a notetaker, and the setting to pick first.

- Published: 2026-11-13
- Publisher: Rig Labs (https://riglabs.ai/)
- Canonical: https://riglabs.ai/newsletter/2026-11-13/

---

Issue 5. Someone still has to turn the notes into a plan after a call. One early user of Claude's new meeting feature [wrote on X](https://x.com/zkevinbai/status/2121026013173858412) that "the work is done before the meeting is over". Our read: the harder choice is which calls it may join, since on a Team plan the default lets it into calls with clients.

In this issue:

1. Claude Meetings: what it adds over a notetaker, and our two bets
2. Gemini 4 Hydrogen: how it compares with Gemini 3.8 Flash and Gemini 4 Argon, and the first job to give it
3. What's new: why an agent can use 40 times more input tokens on one task through an MCP server than through a command-line tool, and side-chat tabs in Codex

#### 01 / The big one

## Will your next meeting end with the plan already written?

Anthropic launched [Claude Meetings](https://claude.com/blog/claude-meetings) on November 12, a beta for organizations that are on a Claude Team or Enterprise plan (Pro and Max do not have it) and use Google Calendar. During the beta, it does not support organizations with zero data retention, a HIPAA-ready configuration or customer-managed keys. Claude joins a Google Meet call as a visible participant, takes the notes, and builds docs, whiteboards and slides that teammates on the invite can edit during the call.

> **Guide:** [Let Claude take your meeting notes and start the work](https://riglabs.ai/guides/2026-11-13-let-claude-take-your-meeting-notes-and-start-the/)
> A step-by-step tutorial: the three team settings, auto-join, and the prompts to use during the call. It comes with a setup file your AI assistant can use to walk you through it.

Many teams already use a notetaker (Granola, Otter and Fireflies.ai are three that Anthropic names as places it plans to send Claude's transcripts). What Claude adds over a notetaker is the doing: a doc started while you talk, and a Slack message sent through a teammate's connected app, with that teammate's permission, per [Anthropic's help page](https://support.claude.com/en/articles/18325154-use-claude-meetings). So what changes, beyond who writes the notes? The bold lines are our forecasts.

- **Teams that use it will leave planning calls with an owner and a date on every action item.** Action items stay vague for a lazy reason: nobody wants to assign them in a recap email after the call. When Claude builds the plan in a doc everyone can see, a task with no owner shows up while the people are still on the call. Anthropic's own example is a team planning an event that says "Claude, start a doc for this plan." Of course a doc made in a few seconds can be wrong in front of everyone. Perhaps that helps: the room fixes a mistake on a shared screen, and ignores one in a recap email.
- **Many Team plan owners will stop Claude from joining calls with outside guests, and set it to internal only, after their first client call with Claude in it.** On Team plans, Claude may join calls with outside guests by default (on Enterprise, an owner must allow it first). With guests allowed, Claude takes notes and posts its notice in the Meet chat, and guests never see the meeting page. Few people want to explain the bot to a client. With internal only, Claude also leaves when another company's notetaker bot joins. So who sets the rules when both companies bring an AI to the call? Decide your setting before the first client call, not during it.

Once your call has started, say this in Google Meet:

```
Claude, start a doc with the plan we have discussed so far, with an owner and a date for each action item.
```

**Your move:** add Claude to one internal planning call this week, not to every meeting, and read the action items on the meeting page before you share them. [Who can use it, what it costs and its limits](https://riglabs.ai/blog/claude-meetings/).

[Give it to your agent](https://riglabs.ai/guides/2026-11-13-let-claude-take-your-meeting-notes-and-start-the/#agent)

#### 02 / Models

## Gemini 4 Hydrogen: a free Gemini 4, and what it costs your app

Google released [Gemini 4 Hydrogen](https://deepmind.google/blog/introducing-gemini-4-hydrogen/) on November 12, the fast, lower-cost member of the Gemini 4 family. It is free in Google Search and on the Gemini API free tier, and in the Gemini app it is in the model drop-down, and free users get it when auto model selection picks it for a prompt.

How it compares: on the paid API, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, twice Gemini 3.8 Flash ($0.75 and $3.75) and 25% below Gemini 4 Carbon and Argon ($2 and $10), per the [pricing page](https://ai.google.dev/gemini-api/docs/pricing). Google says Hydrogen finishes tasks 2.5 times faster than 3.8 Flash at its default thinking level, and scores 51 on the Artificial Analysis Intelligence Index, against 50 for Argon. Simon Willison [calls the Gemini 4 pricing confusing](https://simonwillison.net/2026/Nov/12/llm-gemini/), and the price gap between Hydrogen and Carbon and Argon changes on December 1, when Carbon and Argon double to $4 and $20. If you only use the Gemini app, none of these prices change anything for you.

If a script or automation of yours calls 3.8 Flash, the first thing to do with Hydrogen is to run one real job on both models and compare them. Paste this into your coding agent:

```
Find where this project calls gemini-3.8-flash. Pick one real job, run it once with gemini-3.8-flash and once with gemini-4-hydrogen, and show me both answers, the time each took and the tokens each used. Do not change the code until I say so.
```

**Your move:** keep Hydrogen only where the faster answer saves you a retry or a review, since it costs twice as much as 3.8 Flash, whose price Google lists through December 31, 2026. [Where it runs and the full price table](https://riglabs.ai/blog/gemini-4-hydrogen/).

[Give it to your agent](https://riglabs.ai/blog/gemini-4-hydrogen/#agent)

#### What's new

### MCP or a command line, measured

For: people who give an agent tools through MCP servers or a terminal. Needs: none.

OpenRouter [measured one GitHub task](https://openrouter.ai/blog/insights/mcp-vs-cli-vs-skills/) on Claude Sonnet 5.5: 1,421 input tokens through the gh command line, against 59,417 through GitHub's MCP server with its 46 default tools. Every run answered correctly. OpenRouter notes it tested one server on one task.

For you: when your agent can run commands and the service has a command-line tool, let it use that; if you keep an MCP server, switch on only the tools you need.

### Codex 0.169

For: people who use OpenAI's Codex in the terminal. Needs: a Codex update.

OpenAI's [November 13 release](https://github.com/openai/codex/releases/tag/rust-v0.169.0) adds side-chat tabs to the terminal workspace, which you have to switch on: you can open several tabs, each for a side question apart from the main task, and they are still there after a restart. The release also lets you search older tasks by name, prompt or folder.

For you: you can ask a side question in its own tab without losing the main task.

---

**Which meeting would you never let an AI join?** Tell us which one and why, we would love to hear it. Reply to this email.
