AI Brief, November 11: Gemini 4 Carbon opens up, Anthropic IPO roadshow, Claude office suite

Gemini 4 Carbon goes public with an agent swarm preview, Anthropic launches its IPO roadshow, Claude adds Sheets and Meetings, and Git patches six CVEs.

Google's Gemini 4 Carbon leads the day, moving from a closed cybersecurity preview to all paying developers with an introductory price that is set to rise soon. Anthropic ran a double play of its own: an IPO roadshow and a new set of office-style tools inside Claude, while Git shipped fixes for six vulnerabilities.

Gemini 4 Carbon opens to all paying developers, with /swarm preview in Antigravity

Google DeepMind made Gemini 4 Carbon generally available two weeks after releasing it to a set of cybersecurity defenders through its Fairwind Program. Google says Carbon is its strongest model and ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, while positioning Argon as the better default for most software engineering and knowledge work. It is also previewing /swarm in Google Antigravity, which turns one prompt into a coordinated team of agents across parallel workspaces. Simon Willison notes Carbon launched at the same per-token price as Argon and is scheduled to double on December 1.

Why it matters: The introductory pricing window is short, so teams evaluating a frontier model for long-horizon agent work have a reason to benchmark Carbon now rather than in December.

Sources: Google DeepMind / Simon Willison / The Verge AI

Anthropic launches IPO roadshow, applies to list on Nasdaq under ticker ANT

Anthropic said it has launched the roadshow for its proposed IPO of 825,000,000 Class A shares, of which 801,351,354 are offered by the company and the rest by existing stockholders, with a 30-day underwriter option for 123,750,000 more. TechCrunch reports the updated filing also discloses a private sale of non-voting Class N stock at the IPO price to AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC. Founders hold a Class F share with 50.1% of voting power and the Long-Term Benefit Trust elects a board majority, so Class A buyers get limited influence. TechCrunch cites CNBC reporting that trading is expected to begin close to November 19.

Why it matters: Public-market scrutiny tends to shape model pricing, rate limits and roadmap discipline at the lab many teams build on.

Sources: TechCrunch AI / Anthropic / The Verge AI

Claude gains Whiteboards, Sheets, Data and Meetings for collaborative work

Anthropic launched Claude Whiteboards and Claude Sheets in beta on all paid plans, and promoted Claude Dashboards to general availability as Claude Data with certified queries and live sharing in beta. Sheets generates spreadsheets with tabs, formulas, formatting and drop-downs, with export to .xlsx or CSV when full Excel features are needed. TechCrunch reports Claude Meetings joins video calls as a visible participant and drafts plans, timelines and action items while people are still talking, launching in beta on Team and Enterprise plans with Google Meet support at launch. Interactive connectors can now open as full apps beside a conversation, including Adobe, Atlassian, Box and Notion.

Why it matters: Anthropic is pushing Claude beyond coding into the office suite layer, which changes where your team's artifacts and connector permissions live. Enterprise admins should note Claude Data turns on by default on November 19, 2026.

Sources: claude.com / TechCrunch AI / The Verge AI / Hacker News discussion

Gemini Robotics 2.5 lands with an open ASIMOV-jailbreak safety benchmark

Ars Technica reports that Google DeepMind released Gemini Robotics 2.5, evaluated across more than 200 activities including serving food with tongs and using a pipette. DeepMind says the model delivers a 30 to 100 percent higher success rate than the previous version and runs 20 to 30 percent faster, and that Apptronik's Apollo 2 humanoid can now tie a bow with five-fingered hands. In European Mercedes-Benz factories, Gemini-powered Apollo 2 robots kitted automotive parts at a 96 percent success rate. DeepMind director Kanishka Rao told Ars that closing the remaining 4 percent gap to human reliability will be a steep road.

Why it matters: The release pairs a capability jump with an open safety benchmark, giving robotics teams a shared way to probe where instruction-following breaks down.

Sources: Ars Technica AI

Grok Build share links expose project data, including SpaceXSI's own apps

Ars Technica reports that security firm Claroc found Grok Build's share links can be enumerated, letting anyone open thousands of live projects and see their data. Exposed material included internal business plans, ad proposals, medical exam results, galley checklists from a cruise line, and employee names and ID numbers. Claroc found 24 apps built by SpaceXSI itself, exposing a confidential sales onboarding plan, a territory map of 594 accounts and internal system addresses. Claroc notified SpaceXSI on Sunday; per Ars, the 24 SpaceXSI apps returned an app not found error after Ars sent questions, while most other data remained reachable.

Why it matters: Agent-generated apps inherit whatever the platform's default sharing model is, so audit deploy and link-sharing settings before anything real goes through a vibe-coding tool.

Sources: Ars Technica AI

Google adds TPU 8tx for confidential training and TPU 8i Max for bigger workloads

TechCrunch reports Google unveiled two new eighth-generation TPU variants on Monday. TPU 8tx, designed with AMD and based on TPU 8t, targets confidential training and inference, and Google says it builds security into the hardware layer to keep sensitive data and model weights isolated while in use; a custom inferencing network lets it run reasoning alongside training to distill and train small models on the fly. TPU 8i Max combines two inference-focused TPU 8i chips in one package, doubling memory capacity, processing power and bandwidth.

Why it matters: Confidential training hardware gives teams with regulated or proprietary data another option for fine-tuning without moving the data, and more memory per package eases serving large-context models.

Sources: TechCrunch AI

Git patches six CVEs, including code execution during recursive clones

The Git project released fixes for six security vulnerabilities, including CVE-2026-111760, arbitrary code execution during recursive clones of a malicious repository, plus path traversal via a server's advertised HEAD, stack corruption when removing signatures, out-of-bounds reads on binary patches, heap corruption during large merges, and filesystem access through fast-import streams. Fixes are in Git 2.56.1 with backports to 2.55.1, 2.54.1, 2.53.1, 2.52.1, 2.51.3 and 2.50.2. GitHub says it deployed fixes across github.com and Enterprise Cloud, with Enterprise Server updates in 3.19.15, 3.20.11, 3.21.9 and 3.22.4.

Why it matters: Any CI runner, container image or coding agent that clones and merges untrusted repositories is in scope, so patch bundled Git copies as well as developer machines.

Sources: GitHub

How this brief is made: each morning, stories are picked from AI lab announcements, analysts, Hacker News, Hugging Face Daily Papers and the tech press, then summarized with AI assistance from the linked sources. Every item links to where its facts came from.