Anthropic put Claude inside your video calls. Google released a new Gemini 4 model built for speed rather than depth. And Hark’s computer-use agent just got meaningfully faster at the web tasks people actually hand it. Three things worth knowing.
Claude Meetings Is Now in Beta for Team and Enterprise Plans

Image: Anthropic
Anthropic launched Claude Meetings in beta on November 12. Claude joins Google Meet calls as a visible participant, posts a notice in the call’s chat, follows the conversation, and starts on work during the meeting. It takes notes, writes action items, and builds docs, whiteboards, and other artifacts live while the meeting is happening. Teammates on the invite can refine what Claude builds while the call goes on.
Claude uses each person’s connected tools with that person’s own permission. It asks the first time, and a yes applies going forward. When someone asks out loud, though, Claude for now asks each time to confirm who is speaking. An answer posted in the call’s chat is visible to everyone on the call.
After the call, every note, transcript, and artifact lives on a dedicated meeting page in claude.ai. Org members on the invite can see it, edit it, and share it within the organization. Claude saves a transcript, summary, and screen descriptions. It does not save audio.
The default behavior around outside guests is conservative. Claude stays out of calls where someone from outside your organization is invited, and leaves if one joins mid-call. An admin can override this, but Claude posts a notice and does not reply in chat on those calls.
On Team plans, Claude Meetings is on by default and an admin can disable it. On Enterprise plans, an admin turns it on for the whole org or a subset. It is not available to organizations with zero data retention, HIPAA-ready configuration, or customer-managed encryption keys during the beta.
Google Meet only for now. Other platforms are coming. Transcript partners include Gong, Attention, Granola, Fireflies.ai, Otter, and Plaud, so the meeting transcript will be sendable to tools teams already use. Notes, transcripts, and in-call answers are free on Team and Enterprise plans during the beta, unless an answer uses one of your connected tools, which counts toward your plan’s usage. Docs, whiteboards, and artifacts built in a call are free through February 10, 2027.
Why it matters: The UX question here is whether Claude in a meeting changes what happens in the meeting, or just automates the cleanup after it. The answer is both. Having a doc appear while the team is still talking shifts how decisions get finalized. Anthropic already launched Claude Sheets, Whiteboards, and Data on November 10 as a workspace toolset; Meetings extends that push into the synchronous part of work. The “Claude stays out by default” setting is the right call. No one wants to discover halfway through a client call that an AI has been taking notes without anyone announcing it. That outside-guest behavior tells you who this is built for: internal teams with established workflows.
Find out more: Anthropic’s Introducing Claude Meetings
Google Releases Gemini 4 Hydrogen: Faster Execution, 1M-Token Context

Image: Google
Google released Gemini 4 Hydrogen on November 12. Where Gemini 4 Carbon and Gemini 4 Argon are built for deep, long-horizon reasoning, Hydrogen is built for speed and execution across everyday coding and agentic work. It is priced at $1.50 per million input tokens and $7.50 per million output tokens, flat across the entire 1 million-token context window.
On the Artificial Analysis Intelligence Index, Hydrogen surpasses Gemini 4 Argon at the default high thinking level (51 vs. 50). At the new xhigh thinking level, it reaches 54. Compared to Gemini 3.8 Flash, it finishes tasks 2.5 times faster at the high default setting and more than triples Terminal-Bench 4.1 scores (56.4% vs. 18.2%) (Source). Google says it cuts fabricated claims by more than 5x compared to 3.8 Flash.
Hydrogen has four thinking levels: low, medium, high (the default), and a new xhigh tier. Google says the levels scale in a smooth, predictable progression, so you only wait and pay for the reasoning depth a task requires.
For agentic workflows, Google is introducing Orchestrator in Antigravity. Orchestrator automatically switches between models like Carbon and Hydrogen within a single session. On Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while consuming 24% less quota and completing 19% faster. The idea is multi-model routing without manual selection.
Google puts Hydrogen at #5 on Agent Arena, the live benchmark at arena.ai, and says it leads its class on HLE-Diamond at 58.8%. On agentic safety, it reduces overreach in tasks by up to 81% and achieves a 94% honesty rate on Google’s internal agent honesty evaluation.
Hydrogen is available today in AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and for paid Google AI Pro and Ultra subscribers in Antigravity.
Why it matters: Flat pricing across the full 1 million-token context window at $1.50 input / $7.50 output is straightforward. No graduated costs as the context grows, which matters for teams building agents over large documents or codebases. The more interesting piece is Orchestrator. Automatically routing between a frontier model and a faster execution model in the same session is the kind of plumbing I would rather not build myself. For more context on where Hydrogen fits in Google’s broader AI push in 2026, that recap covers the platform setup. Whether Orchestrator delivers in practice is a question for teams who build on it. Google’s own benchmarks are promising. The external ones are more credible.
Find out more: Google DeepMind’s Introducing Gemini 4 Hydrogen
Hark’s Handoff² Computer-Use Agent Is 2x Faster with an 82% Task Success Rate

Image: Hark
Hark launched Handoff² on November 12, rolling it out to all Hark Pro users. Handoff is Hark’s computer-use agent: it handles web tasks on your behalf so you do not have to open the tab yourself. Book a flight, cancel a subscription, fill out a warranty form, apply for a job. Hark says Handoff has handled over $20 million in purchases and logged over 5.6 million browsing sessions since the product launched in October.
Handoff² averages 2x the speed of the original. The model itself is 46% faster, and additional backend improvements add further gains. Hark says it has seen Handoff² run up to 3.8 times faster on flight booking specifically, and more than 5 times faster on long research tasks.
On Hark’s internal browsing benchmark, 478 tasks across 238 websites, Handoff² completes 82% of tasks, up from 73% for v1, while taking 46% less time per task. Near-frontier accuracy is their benchmark framing: within 5 points of Opus 5.5 at maximum settings. On their internal alignment benchmark, covering behaviors from never checking out without approval to asking for clarification only when a request is truly ambiguous, Handoff² scores 93.8%, up from v1’s 90.3%.
The speed gains come from the training approach. Hark trained Handoff² with reinforcement learning penalties that push the agent to accomplish tasks in fewer turns, batch form-field actions together, and recognize repeatable action patterns mid-task to reuse them as composites. As training progresses, the model takes more actions per turn and needs fewer turns total. Tasks with more browsing steps benefit more from this.
Why it matters: A computer-use agent at 82% task completion on its maker’s internal benchmark is a practical number. 73% for v1 meant roughly one in four tasks went unfinished. Getting that to roughly one in five is a real change in reliability for daily use. The speed improvements matter more than they look in isolation. Hark says it has seen many instances where a 10-minute task shrinks to 3 minutes and a 45-minute task drops to 10. Hark uses Claude Opus 5.5 as their near-frontier accuracy reference point. For flight booking and scheduling, speed is half the value. Waiting ten minutes for an agent to finish something you could have done in five is a UX problem. Handoff² addresses that directly.
Find out more: Hark’s Introducing Hark Handoff²
