Google DeepMind

Google DeepMind, in partnership with Artificial Analysis, has launched the Multimodal Knowledge Agent (MKA) benchmark to evaluate agentic knowledge work over complex visual files. It comprises 200 private real-world tasks across 10 industries, requiring agents to produce deliverables such as slide decks, spreadsheets, and reports from dense visual inputs like architectural drawings, engineering schematics, and clinical scans, emphasizing realistic visual complexity and domain specificity.

Google released Gemini Omni Flash 1.2 on November 5, taking second place on Artificial Analysis' AA-Video-Editing v1.1 leaderboard while climbing from #12 to #6 in text-to-video and #17 to #14 in image-to-video. The model supports text-to-video, image-to-video and video editing with native audio and clips up to 15 seconds, priced at $6.26 per minute of 1080p video, 31% below its predecessor, and will replace version 1.1 on the Gemini API on December 3.

Arena.ai's Agent Arena analysis finds that Gemini 4 Carbon becomes increasingly cost-competitive as task complexity rises. On typical tasks, GPT-6 Astra costs about 3.4x and Claude Opus 5.5 about 3.9x as much as Carbon, while across the most expensive 5% of sessions Opus rises to 4.2x and Astra's cost roughly doubles relative to Carbon. The finding is relevant for budgeting long-horizon agent workloads.

Google DeepMind and Artificial Analysis have partnered to launch MKA (Multimodal Knowledge Agent), a benchmark evaluating how well agents handle visually complex tasks such as cross-referencing blueprints and reading scientific charts, which existing evaluations do not cover. Gemini 4 Carbon ranks second on the benchmark at 6x lower average cost per task than the top model.

Google DeepMind, in partnership with Artificial Analysis, launched the Multimodal Knowledge Agent (MKA) evaluation, a benchmark of 200 private real-world visual tasks spanning 10 industries. Agents must interpret dense visual files such as architectural drawings, engineering schematics and clinical scans, then produce deliverables including slide decks, spreadsheets and reports. Scoring uses expert rubrics split into Climb and Regression criteria, averaging about 70 weighted items per task, with Claude Sonnet 5.5 leading in several domains.

Google DeepMind unveiled Gemini Robotics 2.5, a vision-language-action model it says outperforms its own prior baselines by 30 to 100 percent depending on task type. On active Mercedes-Benz production lines in Hungary and Germany, Apptronik's humanoid Apollo 2 sorted metal parts with a reported 96 percent success rate. DeepMind also uses Gemini 4 Argon as a simulation engine for training and open-sourced the ASIMOV-Jailbreak benchmark to test robotics model safety.

Google has made Gemini 4 Carbon generally available to developers, enterprises, scientists, and consumers, positioning it as its strongest frontier model for cybersecurity capabilities and long-horizon agentic workflows. Carbon sits above Gemini 4 Argon and targets extreme tasks such as complex multi-step coding, debugging, and open-ended scientific research, while Argon remains the default for most software engineering and knowledge work. Carbon was previously released to select cybersecurity defenders through the Fairwind Program.

Google has made Gemini 4 Carbon generally available to developers, enterprises, scientists, and consumers, positioning it as its strongest frontier model with leading performance in cybersecurity and long-horizon agentic workflows. Previously released to defenders via the Fairwind Program, Carbon targets complex multi-step coding, debugging, and open-ended scientific discovery, while Gemini 4 Argon remains the default for most software engineering and knowledge work.

Google DeepMind used teams of Gemini Carbon agents to produce roughly 6.5 million lines of Lean proofs showing that Signal's actual compiled code keeps messages private, verified down to the processor instruction level and checked independently by Lean so no trust in the AI is required. The effort took about four weeks versus years for expert teams, with humans setting security goals and reviewing the ~1% flagged by agents. It is claimed as the first end-to-end formal verification of a widely used protocol at the machine-code level.

Google DeepMind has released Gemini Robotics 2.5, a vision-language-action model with substantially improved dexterous manipulation that has already performed real work in Mercedes-Benz factories and enables Apptronik's Apollo 2 humanoid to tie a bow with five-fingered hands. The team also introduced ASIMOV-jailbreak, an open safety benchmark aimed at evaluating robots operating alongside humans.