Gemini 4 Carbon opens up and Claude adds whiteboards and spreadsheets
Gemini 4 Carbon opens to paying developers, and Claude brings shared whiteboards and spreadsheets into its workspace. This issue also covers open image model weights, offline speech, a robotics pilot on production lines, and Anthropic’s IPO roadshow.
01
Models and products
7 stories
2026-11-10Google DeepMind
Gemini 4 Carbon opens to paying developers
Google DeepMind opened paid API access to Gemini 4 Carbon for developers and enterprises through Google AI Studio and Gemini for business. It is also rolling out the model to Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search.
The public API model costs $2 per million input tokens and $10 per million output tokens during the introductory period, which ends November 30. The rates then rise to $4 and $20 respectively. In one demonstration, a group of agents converted seven public planning PDFs into an interactive 3D model of a building.
2026-11-11MiniMax
MiniMax releases H3-Image weights
MiniMax released the weights for H3-Image, an image model built on H3. It can generate images from text or reference images, edit existing images, and accept up to 16 references. MiniMax says it can produce a 4,096 × 4,096 image in under 10 seconds.
The weights are available on Hugging Face. Hosted access is available through the MiniMax API, MiniMax Code, and Hailuo AI. During launch week, generation requests at 4K or higher cost half the standard rate.
2026-11-10Anthropic
Claude adds shared whiteboards and spreadsheets
Anthropic launched beta versions of Claude Whiteboards and Claude Sheets. Teams can organize notes, timelines, and owners on a shared board, or collaborate on spreadsheets Claude builds with formulas, drop-downs, and multiple tabs. Sheets export to XLSX or CSV.
Claude Dashboards is now generally available as Claude Data, with certified queries and live sharing still in beta. All three products are available on paid plans. Team plans enable them by default. Enterprise plans leave Whiteboards and Sheets off by default and will turn Claude Data on by default on November 19.
2026-11-10Anthropic
Claude opens full third-party apps alongside chat
Claude expanded its interactive connectors so supported apps can open alongside chat or in full screen. Users can edit a Notion or Confluence page in their own account, ask Claude to make a change through the connector, then approve the action and see the result in the same workspace.
The update builds on MCP Apps and extends support for app tools and embedded sites. Developers can build with the capabilities in public preview. Enterprise administrators can restrict connector tools using the existing settings.
2026-11-10n8n Labs
n8n Labs previews Nod, a local personal agent
n8n Labs released a research preview of Nod, an open-source personal agent for Mac and Windows. It runs its own browser in a local virtual machine, where it can use connected apps and regular websites without an n8n account or server.
Users can choose a hosted model provider or connect a local model. Nod asks before sending, paying, posting, or deleting anything beyond the user’s request. That check is itself model-based. Local tasks stop when the host goes to sleep.
2026-11-10Adobe
Adobe Firefly adds Quick Actions and previews reusable workflows
Adobe announced general availability of Quick Actions in Firefly at MAX, with ready-made workflows for tasks such as creating video thumbnails. Image tools for editing text, retouching, and relighting are also generally available.
Firefly Graph is in beta and lets creators assemble their own repeatable workflows. Firefly Builder, a separate tool for prototyping websites and apps, is in private beta.
2026-11-10Google Developers
Google AI Edge Gallery adds offline dictation and read-aloud
Google AI Edge Gallery added on-device speech recognition and text-to-speech. AI Chat can transcribe speech as it arrives, pass the text to a small local model, and read the response aloud on the device. Google showed the offline workflow with Gemini Nano and Gemma 4.
Android developers can use system speech models through ML Kit. Google says its recognizer covers more than 100 languages. For desktops, browsers, and other platforms, it also offers MediaPipe Tasks interfaces and open-weight speech bundles that run on LiteRT.
02
Engineering and research
7 stories
2026-11-10Google DeepMind
Gemini Robotics 2.5 sorts parts in a Mercedes-Benz pilot
Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model. It worked with Apptronik and Mercedes-Benz to test Apollo 2 humanoid robots on active production lines in Hungary and Germany.
The robots picked metal automotive brackets from supply bins and placed them in containers used to sequence parts for assembly. DeepMind reported a 96% success rate for that pick-and-place task in the proof of concept.
2026-11-09NVIDIA
NVIDIA introduces GR00T 2 and Isaac Lab 3.0 at CoRL
NVIDIA introduced GR00T 2, an open world action model, at CoRL. In NVIDIA’s evaluation, it completed 40 of 50 tasks at least once without task-specific demonstrations. Average success across those tasks reached 71% with one demonstration per task and 95% with five.
The accompanying Isaac Lab 3.0 release lets researchers switch between NVIDIA PhysX and the Newton physics solver. Newton can simulate deformable objects such as cables and cloth so robots can practice manipulating them.
2026-11-10NVIDIA
NVIDIA donates OpenShell to the CNCF
NVIDIA announced at KubeCon that it is donating OpenShell, its open-source agent runtime, to the Cloud Native Computing Foundation. OpenShell enters as a CNCF Sandbox project, with organizations using it able to participate in its governance.
OpenShell enforces deterministic policies as agents call tools, access inference services, and run workloads. Its extension points let organizations connect their own compute, inference providers, credential stores, and monitoring tools.
2026-11-10GitHub
Copilot CLI automates C++ language server setup for MSBuild
The Microsoft C++ Language Server for GitHub Copilot CLI now bundles an MSBuild compile-command extractor. When a project opens, it builds the compilation database needed for features such as go-to-definition and find references.
Users need the language server plugin and an active Copilot subscription. CMake projects can still use an existing skill to generate compile commands. Custom build systems can supply a compilation database or use their own extractor.
2026-11-10GitHub
GitHub will pause Dependabot alerts on inactive repositories
Starting December 8, GitHub will pause Dependabot alerts on most repositories that have been inactive for a year. Existing alerts are archived. A push to any branch or tag, or manually turning alerts back on, restores the alerts and restarts the one-year clock.
GitHub skips repositories whose security configuration requires alerts and enterprises that prohibit repository administrators from toggling them. The change covers GitHub.com and GitHub Enterprise Cloud with data residency. GitHub Enterprise Server is outside its scope.
2026-11-10OpenRouter
OpenRouter compares 17 OCR approaches on eight scanned pages
OpenRouter tested 15 vision models and Mistral OCR on eight simulated scanned pages. Each approach checked 88 key strings three times, for 264 checks, with a single Tesseract run as a reference. Claude Opus 5.5 got 262 checks right at an estimated $16.42 per thousand pages. Mistral OCR got 258 right at an estimated $2.02.
The team also compared pairs of lower-cost models. Claude Haiku 5.5 made 35 errors across 2,964 words. DeepSeek V4.1 Flash disagreed on 34 of those errors and flagged 10 correct words. The pair cost an estimated $0.62 per thousand pages.
2026-11-10Google DeepMind
DeepMind uses AlphaProof Code to verify Signal’s core cryptography
Google DeepMind used AlphaProof Code to formally verify the core cryptographic implementation of the Signal protocol. Researchers worked with models to express security goals as mathematical specifications. Experts reviewed those specifications and Lean checked the resulting proofs.
DeepMind says the verification took four weeks and reached down to compiled machine instructions. The result establishes that the core code in scope meets the security properties specified by the experts.
03
Industry
1 story
2026-11-10Anthropic
Anthropic starts IPO roadshow with a $70–$78 price range
Anthropic launched the roadshow for its proposed initial public offering of 825 million Class A shares. It expects to price the shares at $70 to $78 each and has applied to list on the Nasdaq Global Select Market under the ticker ANT.
Anthropic plans to sell 801,351,354 shares, while existing holders are offering 23,648,646 shares and will receive the proceeds from those sales. The company said its SEC registration statement had yet to take effect at the time of the announcement.