The AI week at a glance
The key verified events from our daily editions, preserving their original event dates.
Google Unveils Gemini 4 Carbon for Cybersecurity
Google DeepMind announced Gemini 4 Carbon on October 28, 2026, describing it as the company's strongest AI model to date. It sits above Gemini 4 Argon, which launched two weeks earlier. On version 5.0 of the Artificial Analysis Intelligence Index, Carbon ranks first with an overall score of 59. The system targets cybersecurity defense workflows. It scans software for vulnerabilities, validates exploitation risks, and writes fixes for underlying code flaws. On the CWE-bench evaluation for root-cause patching, Carbon achieved a 71.0 percent success rate. Google is releasing the model in phases, starting with a subset of cyber defense teams while testing guardrails with the United States government. General access for developers and enterprises will follow later. API pricing starts at $2 per million input tokens.
Security teams can locate and repair subtle software defects much faster. This reduces the risk of malicious exploits across large codebases.
Mistral Releases Mistral Large 4 Open Weights
On October 29, 2026, Mistral released the open weights for Mistral Large 4, nicknamed Le Chonk. The multimodal model contains one trillion parameters. It processes both text and images while offering three reasoning modes to balance performance and inference costs. According to Mistral, the system scores 51 percent on the Artificial Analysis Cyber Index, which tests finding and patching vulnerabilities. It also reports a 96.6 percent refusal rate on malicious cybersecurity requests. Developers can host the model on local infrastructure or access it through cloud endpoints.
Teams gain access to a frontier-class model with downloadable weights. Organizations can inspect code and automate defensive security entirely on premises without sending proprietary data to third parties.
GitHub Launches Persistent Copilot Memory on Paid Plans
GitHub announced GitHub Intelligence on October 28, 2026, making Copilot Memory generally available on paid plans. Copilot now remembers developer preferences and project conventions across different sessions. In preview reviews, sessions using memory showed 93 percent better adherence to developer preferences, eliminating the need to retype project instructions.
Software teams save setup time because their AI coding tools retain repo-specific guidelines across daily development tasks.
Google Cloud Releases Gemini Robotics ER 2 in Public Preview
Google Cloud launched Gemini Robotics Embodied Reasoning 2 into public preview on October 27, 2026. The model acts as a high-level cognitive brain for physical machines. It processes live video streams, plans tasks over several minutes, and translates language prompts into actions for robotic arms and humanoid hardware.
Enterprises can automate complex physical workflows without having to program every single motor step manually.
Google DeepMind Announces Genie 4 for Real-Time Simulation
Google DeepMind announced its Genie 4 world model on October 27, 2026. A world model creates an interactive, simulated environment from text or images much like a video game engine. Genie 4 generates native audio on the fly, matching background sounds, music, and spoken dialogue. Users can steer active scenes in real time with text prompts. They can shift the camera, alter the weather, or introduce new characters. The model can maintain a coherent scene for up to two minutes. Genie 4 is currently rolling out to trusted testers. Google plans to release it later to Google AI Ultra subscribers and API partners.
Simulating dynamic environments lets teams train robots safely before real deployments. It also provides faster prototyping tools for games and films.