Google has made Gemini 4 Carbon available, targeting complex reasoning and long-horizon tasks in software engineering, cybersecurity, and scientific research. Developers can access it through the paid API in Google AI Studio, while Google AI Ultra subscribers can also use it in the Gemini app and Antigravity. Introductory API pricing is $2 per million input tokens and $10 per million output tokens; starting December 1, the rates will rise to $4 and $20, respectively. Antigravity is also introducing the multi-agent /swarm command, which can coordinate different models to divide and execute tasks.

Stripe has completed its acquisition of OpenRouter, an AI model access platform, about three months after reaching an agreement to acquire it. OpenRouter provides more than 10 million developers and businesses with access to multiple model providers and helps users manage model costs and data strategies. The platform says it will remain neutral, allowing models and providers to compete on quality, reliability, and price. Since the acquisition agreement was announced, weekly token usage on the platform has reached 2.3 times its previous level.

MiniMax has released H3-Image, an image-generation model, and open-sourced its weights. The model supports text-to-image generation, image editing, and up to 16 reference images; according to the company, it can generate a 4096 × 4096-pixel image in about 10 seconds. Users can access it via API, MiniMax Code, and Hailuo AI. During the first week after launch, images of 16 megapixels or higher are available at half price. vLLM-Omni has also provided a corresponding Docker image and deployment configurations for different hardware.

NVIDIA has introduced GR00T 2, a robotics model for long-horizon tasks that supports zero-shot deployment through X-Adapt. The release includes the model weights, framework, and 100,000 hours of training data for robotics developers.

Google DeepMind has released Gemini Robotics 2.5, a vision-language-action model that can handle more than 200 tasks, including grasping food, tying bows, and pipetting. In evaluations, its task success rates were 30% to 100% higher than the baseline, while completion speed increased by 20% to 30%. In industrial tests with Apptronik and Mercedes-Benz, the Apollo 2 robot achieved a 96% success rate in autonomously sorting automotive parts. The model also supports reasoning-guided behavior when it encounters obstacles or people.

Anthropic has introduced Claude Whiteboards, Sheets, and Data for planning, creating tracking sheets, and organizing data, respectively, as well as Claude Meetings, which can join video meetings and capture follow-up items. New interactive connectors let users open third-party tools such as Adobe, Atlassian, and Box in Claude, edit content themselves, or ask Claude to help with tasks. The connector is rolling out to the web and desktop.

OpenAI has begun testing business features in ChatGPT. Eligible companies can claim a profile page, manage business information, and view insights into user interactions. Businesses can also choose to create a Business Agent to answer customer questions and help introduce products or services. Business Hub lets verified businesses update their information and add product data. Conversations between customers and business agents will identify product information supplied by the business, distinguishing it from independent answers from ChatGPT.

Google DeepMind has introduced AlphaProof Code, a formal verification framework that combines human expertise with AI agents to mathematically verify key security properties of the Signal Protocol implementation. The verification targeted compiled code rather than just the protocol design, generated more than 6 million lines of Lean code, and took four weeks. To test the approach, the system also identified two vulnerabilities that had already been fixed.

Ant Lingbo has open-sourced its tactile foundation model, LingBot-Touch. The model converts signals from different sensor types, including magnetic, vision-based, and piezoresistive sensors, into a unified tactile representation for reuse by robot action models. It was pretrained on 4,983 hours of real-world tactile data, covering about 464,000 interaction records and 1.33 billion frames. Across 26 real-robot tasks, integrating the model increased the average task success rates of four action models by about 14 to 26 percentage points.

Meituan has officially open-sourced the LongCat-2.5 model weights under the MIT License, with downloadable files totaling about 3.56 TB. The model has 1.6 trillion total parameters, with about 48 billion activated per token, and supports native visual understanding. It is designed for long-horizon agent tasks such as developing complex codebases, building projects, and office work. The team also optimized long-context training and W4A4 inference for domestic computing hardware.