Issue W47/25: OpenAI Launches GPT-5.1, Google DeepMind Reveals SIMA 2, and Anthropic Disrupts AI Espionage - November 17, 2025
Last week, OpenAI released the adaptive GPT-5.1 and Codex models, Google DeepMind introduced the self-improving SIMA 2 agent, and Anthropic exposed a state-sponsored cyberattack orchestrated by AI.
General News
OpenAI began piloting group chats for ChatGPT in select APAC regions, while Anthropic revealed it disrupted a state-sponsored cyber-espionage campaign where the AI agent Claude was used to orchestrate attacks. In autonomous driving, Waymo expanded its freeway operations to connect San Francisco and San Jose. Meanwhile, Cursor announced a $2.3B Series D funding round, signaling strong investor confidence in AI coding agents.
Google DeepMind
Google DeepMind introduced SIMA 2, a generalist agent capable of self-improvement and navigating complex 3D environments like those generated by Genie 3. On mobile, the Pixel feature drop includes "Nano Banana", a Gemini-based image generation model, while Veo 3.1 was updated to support multiple reference images for better video control. Google also rolled out significant updates to Gemini Live, focusing on faster turn-taking and more expressive voice interactions.
Foundation Labs & Model Updates
OpenAI released GPT-5.1 and GPT-5.1-Codex, featuring "adaptive reasoning" that adjusts compute based on task difficulty, though it launched without detailed benchmark reports. Baidu launched ERNIE-4.5-VL-28B, an Apache 2.0 model claiming SOTA document understanding, while Anthropic rolled out Structured Outputs for Claude to guarantee JSON schema compliance. Additionally, Jan-v2-VL was released, optimizing 8B models for long-horizon agentic tasks.
AI Developer Topics
Coding agent platforms moved quickly to adopt new models, with Windsurf making GPT-5.1 its default and Cursor integrating the Codex variant. LangChain released a Deep Agents podcast focused on context engineering and sandboxed execution environments. Hugging Face announced a major partnership with Google Cloud to optimize open model deployment on Vertex AI and GKE. Furthermore, Nous Research slashed pricing for Hermes 4 by 70% to encourage wider adoption in agentic workflows.
Research News
OpenAI published research on training models with sparse circuits to improve interpretability without sacrificing performance. A new initiative called AELLA aims to process 100 million scientific papers into structured summaries to democratize research access. On the theoretical front, researchers proposed "Looped LLMs", suggesting that iterating a model's own computation can effectively increase depth at inference time. Finally, a new Code Arena benchmark was launched to evaluate how well agents plan and ship full web apps.
Others Topics
NVIDIA launched the RTX Pro 5000 workstation GPU, featuring 72GB of VRAM to support local LLM and VLM workloads. In social trends, a Japanese woman married a "persona" created within ChatGPT, utilizing AR glasses for the ceremony. Meanwhile, controversy erupted around the Perplexity Partner Program as users reported mass bans and allegations of fraud regarding payouts.