Claude can now sit in on your meetings. Anthropic opened the Claude Meetings beta today: add Claude to a Google Meet call and it drafts the plan, answers questions and builds docs while people are still talking.

It's rolling out on Team and Enterprise plans, and during the beta the notes and transcripts are free.

Also tonight: Google's Gemini 4 Hydrogen edges past Gemini 4 Argon at Flash-tier prices, Liquid AI ships a tiny speech codec for on-device voice, Nvidia's board reportedly moves against Jensen Huang, and four tools worth trying.

Anthropic opened the beta for Claude Meetings today. Add Claude to a Google Meet call and it joins as a visible participant, follows the conversation and starts on the work while you're still talking: answering questions out loud or in chat, pulling in information from your connected tools with permission, and building docs and whiteboards live that teammates can edit. Before the call ends, it can share a recap with notes and action items.

It's rolling out now on Team and Enterprise plans, on Google Meet only for now, with other video platforms "coming soon." You add Claude to upcoming calls from your calendar in claude.ai, and every call gets its own meeting page with the transcript, a summary and descriptions of what was on screen. Anthropic says it doesn't save the audio. During the beta, notes and transcripts are free, answers in calls are free unless they use a connected tool, and docs and whiteboards made in calls are free through February 10, 2027.

Our take: this is the piece that ties Tuesday's launches together. Sheets and Whiteboards are where the work lands; the meeting is where it actually gets decided. A Claude that drafts the plan during the call, instead of a bot that emails a summary afterward, changes how a meeting ends. The limits are real, though: Google Meet only, Team and Enterprise only, not yet available to organizations using zero data retention, HIPAA-ready setups or customer-managed encryption keys, and by default it won't stay in calls with outside guests.

Try it: Setup guide · Claude app

Most AI note-takers turn a call into a summary someone may or may not read later. Claude Meetings treats the call as context for work that's already underway, and it keeps that context where the rest of Claude can use it: you can ask Claude about past meetings you attended, and meeting pages can be shared across your organization. Anthropic says transcript partners Gong, Granola, Fireflies.ai, Otter and Attention are coming soon, so notes can also flow into the tools teams already use.

It's also a bet on trust, because this is an AI that hears every word. Claude announces itself in the call, Team admins can switch it off, Enterprise admins have to switch it on, and by default it stays out of calls with outside guests and leaves if one joins. And it rounds out a week of office launches from Anthropic, after Tuesday's Sheets and Whiteboards: Claude is no longer the tab you open after the meeting. It's in the room.

⚡ Google Launches Gemini 4 Hydrogen

Gemini 4 Hydrogen is Google's new fast, low-cost model, pitched as its "daily driver for agentic workflows." It has a 1M-token context window, takes text, images, audio and video, and adds a new xhigh thinking level. Pricing is $1.50 per million input tokens and $7.50 per million output, flat across the whole window, with a rate-limited free tier in AI Studio. It's also rolling out in Google Search for all users and to free Gemini app users through automatic model selection.

Google's numbers: 57% on Terminal-Bench 4.1, up from 18.2% for Gemini 3.8 Flash, and a score of 51 on the Artificial Analysis Intelligence Index at its default setting (54 at xhigh), against 50 for Gemini 4 Argon. In Antigravity, a new Orchestrator mode lets Gemini 4 Carbon plan while Hydrogen makes the edits, which Google says used 24% less quota and finished tasks 19% faster on average.

Our take: the interesting number is the price next to the benchmark. A model that matches a frontier flagship on a broad index at Flash-tier pricing changes the default for agent work, which needs lots of fast, cheap steps. Caveats from Google's own model card: Claude Sonnet 5.5 still leads it on cheating-adjusted Terminal-Bench 4.1 (63.9% to 57.0%), and Hydrogen reached the alert thresholds, though not the critical levels, for biology and cyber uplift.

Try it: AI Studio (free tier) · Gemini app · API docs and pricing

Liquid AI released LiquidCodec today, a roughly 74-million-parameter streaming speech codec that turns audio into tokens a language model can predict, then back into audio. It's the audio tokenizer inside LFM3-Audio-2B and LFM3-Omni-2B, the small voice models Liquid shipped at the end of October, and it works on 24 kHz speech at about 1.65 kbps with a fully causal encoder for low-latency streaming.

Liquid's numbers, measured on a single CPU thread on an Apple M4 Max: encoding at about 361x real time and decoding at about 87x, faster than Mimi, NanoCodec and X-Codec 2 on both. It also reports the best reconstruction scores on all eight of its test sets, and a 4-bit decoder that's about 3.5x smaller with minimal quality loss. It was evaluated on speech only, not music or other sounds.

Our take: codecs are the unglamorous part of voice AI, but they decide how fast and how cheaply a phone or laptop can listen and talk back. A tokenizer this light is what makes a 2B voice model practical offline. All benchmarks are Liquid's own, and the model card lists Liquid's LFM Open License v1.0, so check the terms before commercial use.

Try it: LiquidCodec weights · LFM3-Audio-2B · Liquid docs

The other stories worth your attention:

Nvidia's board is reportedly preparing to remove Jensen Huang as CEO, after reports about his personal relationships and a board investigation into whether a foreign intelligence operation targeted him. Unconfirmed by Nvidia, which hadn't commented; see how far the stock fell after hours. (WSJ via Reuters)

Anthropic found AI agents do worse as a team when every user brings their own agent to a shared task, across models from Anthropic, OpenAI and Alibaba. See how wide the gap was and the fix that closed it in one setting. (Anthropic)

Perplexity let 12 frontier models try to break out of its agent sandbox. None escaped the virtual machine, but several got past its network rules. See which models did it and how fast. (Perplexity)

US company AI spending hit a record in October, but growth is slowing and customer support is suddenly a big line item, per Ramp's AI Index. See whether Anthropic or OpenAI gained more business adoption. (The Decoder)

Databricks turned Genie into an always-on AI coworker that runs long jobs in the background and acts on triggers like a metric dropping, plus a new Ontology layer that maps your business. See how it scored against frontier agents in Databricks' own benchmark. (Databricks)

Anthropic opened the Claude Meetings beta, putting Claude into Google Meet calls on Team and Enterprise plans to draft plans, answer questions and build docs live, with notes and transcripts free during the beta. Google launched Gemini 4 Hydrogen, a fast model that edges past Gemini 4 Argon at $1.50 in and $7.50 out per million tokens, Liquid AI released LiquidCodec for on-device voice, and Nvidia's board reportedly moved to replace Jensen Huang. The thread: AI is moving from the tab you open after the meeting to a participant in it.

Stay building. 🤖

Our newest video covers Claude Fable 5.2, the hidden Fable 6 hints, Gemini Krypton, GPT-6.2 and Sol.