AI Today: Gemini 4 Carbon and Gemini Robotics 2.5
Authorship and approval
Published and updated
State of knowledge as of 2026-11-11. Source material collected on 2026-11-11. Google is expanding Gemini Intelligence to additional devices, publishing the MKA benchmark with 200 tasks, and releasing the Gemini 4 Carbon model. Separately, it introduces Gemini Robotics 2.5, which achieved 96% success in a part-sorting pilot. OpenAI is launching a beta test for company and agent profiles in ChatGPT, with Business Agents access for eligible e-commerce and online service companies in the USA.
Key takeaways today (TL;DR)
- Gemini Intelligence: new features are rolling out to more devices, and task automation is expanding to the UK, India, and Japan.
- MKA: the benchmark covers 200 tasks modeled after work across 10 industries and 83 professions.
- Gemini 4 Carbon: the model took first place in three benchmarks and is available via API for paying developers and enterprises.
- Gemini Robotics 2.5: the model achieved 96% success in a part-sorting pilot on Mercedes-Benz lines.
- Businesses in ChatGPT: companies can manage profiles and offerings, and eligible US businesses can deploy Business Agents.
Google expands Gemini Intelligence to additional devices
Google is rolling out Gemini Intelligence to more devices, including Pixel 10 and motorola signature 27. Users of these phones receive new features, and task automation is coming to additional countries.
- Rambler converts speech into structured text in real time; it works on selected Pixel 10 and motorola signature 27 models.
- Task automation expands to the UK, India, and Japan; Gemini can trigger actions within apps, such as ordering food or booking a table.
- Create My Widget allows you to describe your own widget in plain words; the feature uses information from Gmail, Calendar, and Search.
- Pixel Watch is set to receive Create My Widget in the coming weeks.
- Device Assistance lets you ask Gemini for help with phone settings, such as freeing up storage or checking app location access.
- Android Auto will soon display personalized suggestions using information from apps. For AI builders: Teams developing mobile apps should check whether these features will cover their users and markets. Google does not provide a precise list of languages and countries, pricing, or API access here.
Google and Artificial Analysis launch benchmark with 200 tasks
Google and Artificial Analysis introduced MKA, a benchmark measuring how AI agents handle complex visual files. Teams can use it to evaluate their own models across 200 tasks modeled after professional work.
- 10 industries and 83 professions represent tasks prepared with experts, ranging from architects to biologists and data analysts.
- Each task reflects a 3 to 8-hour project; it includes multimodal files and a rubric with about 70 criteria.
- The rubric weights 75% factual correctness and 25% visual appearance. The automated evaluator achieved 96% agreement with expert ratings.
- All models were tested in the same open Stirrup environment, with the same tools, sandbox, and turn limit.
- Gemini 4 Carbon took second place in quality; on average, it cost six times less than the top-scoring model. It is available through the Gemini API. In the lab: The lab database features a text on designing benchmarks for knowledge work; MKA focuses on tasks requiring the analysis of visual materials. For AI builders: Teams building agents to work with visual files can test them on 10 public MKA tasks. The announcement does not specify the task license, Polish language support, or Gemini API pricing.
Google releases Gemini 4 Carbon, leader of three benchmarks
Google DeepMind has released Gemini 4 Carbon, its most powerful model for cybersecurity, programming, and scientific research. It supports long agentic workflows, meaning tasks executed in stages, and paying developers and enterprises can use it via the API; it took first place in three tests.
- CursorBench 4.0, WANDR, and AutomationBench: Carbon took first place, and the task cost was more than twice as low as the next external model.
- In the multimodal MKA benchmark, evaluating agent performance on charts and images, Carbon took second place; the average task cost was more than six times lower than the leader's.
- The API in Google AI Studio and Gemini for enterprise is available to paying developers and businesses. The model is also coming to Google AI Ultra subscribers in Antigravity, the Gemini app, and AI Mode in Search.
- Introductory pricing: $2 per million input tokens and $10 per million output tokens — the same as for Gemini 4 Argon.
- Preview of the /swarm feature in Google Antigravity turns a single prompt into a coordinated team of agents that can be tracked in real time. In the lab: We wrote about Gemini 4 Argon: Built to solve your hardest problems; DeepMind is splitting model roles: Argon handles most tasks, while Carbon tackles the hardest and multi-step ones. For AI builders: Teams building agents can check task costs and API access. The announcement does not specify usage limits or EU availability.
Google increases robot dexterity; sorting pilot reaches 96%
Google showcased Gemini Robotics 2.5, a vision-language-action model: it combines image and instruction inputs with robot actions; in a part-sorting pilot, it achieved 96% success. The model is designed to teach robots more skills and is being rolled out to Google technology partners.
- Over 200 tasks in tests included feeding food with tweezers, tying a bow, and pipetting.
- Compared to previous models, success rates increased by 30–100% depending on the task; robots performed them 20–30% faster.
- Apollo 2 sorted metal parts into bins on Mercedes-Benz assembly lines in Kecskemét and Sindelfingen. This is a pilot for preparing component kits for car assembly.
- The model works with various robots, including the Apollo 2 humanoid and the Franka Duo platform with two-finger grippers. In the lab: The lab database features an earlier description of Gemini Robotics 2, focused on whole-body control; version 2.5 emphasizes faster skill learning and specific task test results. For AI builders: Robot integrators and manufacturing plants can check if their tasks resemble Google's tests or pilot. The source does not provide pricing, detailed access terms beyond providing the model to partners, or licensing information.
OpenAI adds company and agent profiles to ChatGPT
OpenAI launched the Businesses in ChatGPT beta so companies can manage their information and interact with users. Customers can talk to corporate agents, i.e., AI agents, about offerings.
- Business Hub allows companies to submit a profile and, after verification, edit data and add product catalogs with pricing and availability.
- Shopify allows merchants to connect their store via Business Hub or the Shopify admin. OpenAI Enterprise customers and advertisers can submit their company through their OpenAI representative.
- Analytics include company mentions, website clicks, and conversation topics where the site appears. OpenAI provides them in aggregate once privacy-protecting thresholds are met.
- Business Agents are available to eligible e-commerce and online service companies in the USA. Companies can also sponsor agents; such slots are labeled and separated from ChatGPT responses. In the lab: The post “How AI is changing professional services and data” discussed sponsored agents; now OpenAI is giving companies tools to manage profiles and offerings. For AI builders: E-commerce and online service companies can check if they qualify for Business Agents; others can verify if profile submission is available in their country. The announcement does not state pricing or the timeline for agent availability in Poland or the EU.
Frequently Asked questions (FAQ)
What does the MKA benchmark measure?
MKA covers 200 tasks from 10 industries and 83 professions, prepared with experts. Each task corresponds to a 3 to 8-hour project, includes multimodal files, and features a rubric with about 70 criteria. Tests were conducted in the same Stirrup environment.
How did Gemini 4 Carbon perform in benchmarks?
Gemini 4 Carbon took first place in CursorBench 4.0, WANDR, and AutomationBench. In MKA it placed second, and the average task cost was more than six times lower than the leader's. Paying developers and businesses can access the API.
Who can use Business Agents in ChatGPT?
Business Agents are available to eligible e-commerce and online service companies in the USA. Companies can submit profiles via Business Hub, and after verification, edit data and add product catalogs with pricing and availability.
Sources
- Gemini Intelligence: New features, on more devices — 2026-11-10
- Introducing the Multimodal Knowledge Agent benchmark — 2026-11-10
- Gemini 4 Carbon is now available — 2026-11-10
- Gemini Robotics 2.5 accelerates real-world general dexterity — 2026-11-09
- Introducing Businesses in ChatGPT — 2026-11-10 --- Text prepared automatically by the aipoc-lab (AI) laboratory based on the publicly available sources listed above. Own summaries, not copies. Pre-human-editing version.
Provenance manifest of this post (JSON) — which signals it was built from, what decisions the loops made, and when a human approved it.
Related posts
Planning in Agentic Systems: Theoretical and Operational Frameworks
Explore the theoretical foundations and operational frameworks of modern planning in agentic systems based on 903 lab signals.
AI Today: GPT-6.2 Astra and Gemini
Discover the latest updates on GPT-6.2 Astra, Gemini task mode, and new specialized features for sales and healthcare.
Architecture and Security of Multi-Agent Systems (MAS)
An analysis of multi-agent LLM systems focusing on orchestration frameworks, verification methods, memory management, and security risks.
AI Today: Gemini 4 Hydrogen and AlphaGenome 2
Google releases Gemini 4 Hydrogen and AlphaGenome 2, while Meta introduces user-customizable spaces in the Muse app.