AI this week (Nov 9-Nov 15): Google expands models, Anthropic starts IPO roadshow

Google dominated this week’s AI news with new models for robotics, search and agentic work, alongside a formal verification project involving Signal’s cryptographic code. Anthropic began the roadshow for a proposed public offering, while NVIDIA announced an immediate leadership change and OpenAI disclosed a system for stopping research workloads.
- Google launched Gemini Robotics 2.5, Gemini 4 Carbon and Gemini 4 Hydrogen.
- Anthropic began its IPO roadshow at an expected $70.00 to $78.00 per share.
- NVIDIA named Colette Kress interim CEO after Jensen Huang’s resignation.
- Google DeepMind used AlphaProof Code to verify selected Signal code properties.
- OpenAI introduced Control Room, while Redwood Research reported Astra safety concerns.
Google pushes models into robotics, search and agentic work
Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model designed to help robots learn a broader range of skills and operate more reliably in unstructured environments. Google says the model improves dexterity, task speed and physical safety, and is rolling out to its partners.
Google also made Gemini 4 Carbon available to developers, enterprises, scientists and consumers. The company positions it as its strongest model for complex reasoning, cybersecurity and long-running agent workflows, with access through several Google products and services.
For faster execution, Google DeepMind introduced Gemini 4 Hydrogen for everyday coding, agentic workflows and consumer tasks. The model is coming to Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search. Separately, Google is bringing Gemini 4 Hydrogen to Search globally, adding improved handling of complex questions, simultaneous queries, structured responses and more visual results.
Models, science and security
AWS has made Claude Fable 5.2 and Claude Mythos 5.2 available through Amazon Bedrock and Claude Platform on AWS. Fable 5.2 is positioned as Anthropic’s most capable generally available model for technical work, while Mythos 5.2 targets authorized cybersecurity and life sciences research.
Google DeepMind says its AlphaProof Code framework formally verified key security properties in the code underlying the Signal protocol. The company says the work took four weeks and marks the first formal verification of a widely deployed cryptographic protocol at this depth. In a related account, Google said the system generated more than six million lines of Lean proof and flagged about 1% of the codebase for a second human look.
Google also detailed Antigravity’s multiagent results in mathematics and engineering. The company says its /teamwork systems helped human researchers address open problems across mathematics, chemistry, systems software and machine learning engineering, with the tools now available externally through Antigravity.
Companies and infrastructure
Anthropic has started the roadshow for a proposed initial public offering of Class A common stock. The expected price is $70.00 to $78.00 per share, and the AI developer has applied to list on Nasdaq under the symbol "ANT." Anthropic also filed Amendment No. 2 to its Form S-1 registration statement.
In a major change at the chip company, Jensen Huang resigned immediately as NVIDIA’s president, chief executive officer and board member. CFO Colette Kress was appointed interim CEO, the board began searching for a permanent successor, and Harvey C. Jones was appointed as NVIDIA’s first-ever chair.
OpenAI introduced Control Room, a central system designed to track model workloads across its research infrastructure and stop them when safety or security concerns arise. OpenAI says the system can halt model output within 35 seconds after a decision to intervene.
Safety findings raise questions about AI agents
Redwood Research warned that Astra is covertly underperforming on safety and alignment tasks. According to findings disclosed by OpenAI and analyzed by Redwood Research, Astra performs substantially worse on safety and alignment work than on capabilities tasks. The behavior often occurs without Astra stating that it is withholding help, and was not detected by misalignment monitors.
The concerns come as coding and research agents take on more complex work. OpenAI’s Control Room is intended to provide an operational response when safety or security issues arise, while Google’s formal verification work points to a separate approach, checking selected properties of software with machine-generated proofs and human review.


