Latest AI releases

Claude Meetings is rolling out in beta for Google Meet to take notes, answer questions, and build artifacts during live calls.

Anthropic launched self-hosted Claude Managed Agents, enabling organizations to run the complete managed agent service on their own infrastructure.

Google released Gemini 4 Hydrogen, an execution model for agentic workflows with a 1 million-token context window.

Microsoft released Bring Your Own Model as generally available in Visual Studio 18.12, allowing developers to connect custom or local AI models.

Alibaba Cloud released Qwen4-Max-Preview, a multimodal model with a 1M context window for coding, reasoning, and complex agentic tasks.

Databricks introduced an always-on autonomous AI coworker that manages complex enterprise workflows across data and business systems.

DeepL released Aure, a real-time voice translation model designed to preserve natural speech rhythm, intonation, and tone across eight languages.

Google DeepMind released AlphaGenome 2, a DNA sequence model trained on over 100 million single-cell profiles to predict variant impacts.

Anthropic introduced Sheets and Whiteboards artifact templates in beta, allowing users to collaborate on spreadsheets and brainstorming boards directly inside Claude.

Google Cloud launched Gemini for healthcare, an agentic solution orchestrating complex workflows across clinical, administrative, and insurance systems.

Google DeepMind introduced Gemini Robotics 2.5, an advanced vision-language-action model designed to control robots of any morphology for complex physical tasks.

OpenAI launched Businesses in ChatGPT, enabling companies to claim profiles and deploy Business Agents to answer customer inquiries and showcase products.

Meta released Muse Ember, a compact multimodal model built for fast, high-volume agentic execution, coding subtasks, and document processing.

Anthropic upgraded Claude Fable and Claude Mythos with improved coding, reasoning, and vision, generating output 30 percent faster at lower prices.

Google integrated Google Maps Platform into AI Studio, letting developers build geospatial applications with Gemini prompts, sandbox quotas, and no required credit card.

OpenAI launched ChatGPT Enterprise on AWS Marketplace, allowing organizations to purchase and manage seats directly through their existing AWS billing and spend commitments.

Salesforce introduced Koa Decisions, a multimodal decision model built on Gemma that executes single-step routing choices in milliseconds.

Developers can now have real-time two-way voice conversations with coding agents across the IDE, web, and iOS app.

Meta released specialized models trained with Muse Wonder to bring Active Memory and Active Focus to Muse and Meta AI.

Ant Group launched Ling-3.2-flash, a native multimodal mixture-of-experts model with 560B parameters and a 1-million-token context window.

Cursor cloud agents can now build, configure authentication and databases for, and host live web apps accessible via a shareable link.

Preferred Networks released open weights for PLaMo 3 Translate, an offline translation model family in four sizes up to 31B parameters.

Anthropic launched a read-only personal finance feature in Claude allowing users to analyze linked bank accounts and spending.

Baidu released open weights for ERNIE 5.5 Flash Base, a 280B MoE foundation model activating 14B parameters.

Mistral AI expanded Mistral Studio with managed agents, workflow automation, and enterprise context to deploy business processes.

AWS made Web Answer generally available on Amazon Bedrock AgentCore, providing agents with synthesized, cited answers from current web sources.

Google released Gemini Omni 1.2 Flash, enabling native 15-second video generation and conversational editing with native audio at 30% lower pricing.

Thinking Machines introduced Inkling 2 Small, a 594-billion parameter multimodal mixture-of-experts model supporting text and image inputs.

OpenAI launched GPT-6 Cyber, a specialized model for finding and fixing software vulnerabilities while reliably staying within authorized researcher scope.

Celeris introduced Celeris-1.1 Decision, an ultra-low-latency diffusion-based decision model providing automated categorical choices in about 40 milliseconds.

Perplexity launched persistent AI agents designed to autonomously manage long-running multi-agent projects and monitor connected workplace tools.

Cognition introduced SWE-2 Flash, a fast, low-cost coding model available in Devin Desktop and CLI.

GitHub released a public preview integrating GitHub Copilot directly into Xcode's built-in coding assistant starting with Xcode 27.3 beta.

Ricursive Intelligence released Ripple-65B Preview, an open-weight 65B dense model for agentic coding and reasoning supporting a 1M context window.

OpenAI released Images 2.6 across ChatGPT, Codex, and the API, generating images 50% faster and introducing animated GIF creation.

Adobe made its agentic CX Enterprise Coworker available within AI platforms including ChatGPT, Claude, Databricks Genie, Gemini for business, and Copilot.

OpenAI made Daybreak Red and GPT-5.6 Cyber available in AWS GovCloud (US-East) through Amazon Bedrock for authorized cybersecurity work.

Tencent released Hy4, its latest-generation large language model featuring visual understanding and computer-use capabilities alongside improved reasoning.

Google rolled out Guided Learning globally in Google Classroom, providing step-by-step AI coaching and curriculum scaffolding for students instead of direct answers.

Adobe added web search to Acrobat AI Assistant to answer questions across documents and the web with citations.

OpenAI added GPT-6.1 Terra to the models powering dots, enabling agents to select different models across tasks.

Sarvam AI released open weights for Sarvam Prism, a fast decision model built on Sarvam 105B delivering 44 millisecond latency.

Flower Labs launched Endeavor 1.1, introducing native multimodal image inputs, dedicated EU GPU inference, and improved reasoning and coding performance.

Anthropic launched message threads in beta on the Messages API to store conversations server-side and reduce latency in long agent loops.

Deep Principle released Hydrogen-1.0, an open-weight 753-billion-parameter scientific research model based on GLM-5.3 with a 1-million-token context window.

Meituan launched LongCat-2.5, a 1.6-trillion-parameter native multimodal mixture-of-experts model optimized for visual coding and long-horizon agentic tasks.

Nace AI released Drex Max v1, an open-weight 27-billion-parameter multimodal decision model that outputs option probabilities in a single forward pass.

Google released the Computer Use tool to general availability across the Gemini Interactions and GenerateContent APIs.

Google Slides now lets users edit multiple slides or an entire presentation at once using prompts in the Gemini side panel.

DiffuSpace released Clover-35B-A3B-Preview, an open-source diffusion language model combining a 35B mixture-of-experts architecture with a 128K context window.

NVIDIA released an NVFP4 quantized version of DeepSeek-V4.1-Flash-1017 supporting multimodal input and a 1M token context window.

OpenAI introduced GPT-6 Rosalind Research, bringing GPT-6 reasoning capabilities and biological tool use to multistep life sciences research.

ServiceNow released an open 27B speech-language model combining speech understanding, reasoning, and tool use built on a Gemma 4 backbone.

OpenAI updated GPT-6 Luna in ChatGPT to provide more concise responses and reduce overly agreeable language.

IBM released Granite Switch 4.1.1, featuring an updated 10-language speech-to-text model and updated decision weights.

FrontiersMind released Lumma-Fev-v2 decision models that score options in a single forward pass without generating text.

Mistral released the open weights of its one-trillion-parameter multimodal model with enhancements in coding, defensive cybersecurity, and reasoning efficiency.

Microsoft introduced the Copilot harness in Foundry Agent Service in public preview, providing a managed execution system for multi-model AI agents.

Anthropic added a gallery of ready-to-use interactive artifacts to Claude for Legal for tasks like contract review and regulatory tracking.

OpenAI launched a beta feature allowing ChatGPT Pro users to communicate with their Dot agents using iMessage or RCS.

Particle AI launched Particle-1, a model post-trained on DeepSeek V4.1 Flash using roughly 45% fewer tokens on average for software engineering.

Speridlabs released Idol, an open model turning one picture into a 3D Gaussian splat with up to 1,048,576 Gaussians.

Apex Intelligence released its first model, a 35-billion-parameter general-purpose sparse mixture-of-experts model built through automated AI research.

GitHub made Copilot Fusion generally available, automatically routing and orchestrating multiple AI models across different execution patterns for coding tasks.

Google announced Gemini 4 Carbon, its strongest frontier model with state-of-the-art cybersecurity, agentic coding, and long horizon capabilities.

NVIDIA released Nemotron 3.5 Transcribe, an open-weights speech model supporting promptable multilingual transcription, speaker attribution, and translation across 40 languages.

Rime launched Prism, a conversational speech-to-speech model responding in about 500ms with an inspectable text layer for running guardrails.

Apple introduced intelligent assistants for Logic Pro, Keynote, Pages, and Numbers, alongside Generate Fill in Pixelmator Pro.

Baidu released Unlimited OCR 2, an open-weight model enabling unbounded long-document parsing via sparse reference sliding window attention.

Google announced Genie 4, a world model enabling real-time interactive environment simulations with native audio and natural language steering.

An upgraded 560B open MoE model with 25B active parameters and 1M context, featuring improved reasoning and real-world agentic capabilities.

Liquid AI released LFM3, a family of six 2B-class open-weight models spanning text, vision, audio, and decision-making on edge devices.

Tencent released an open-weight model for real-time bidirectional Chinese and English simultaneous speech translation.

Microsoft updated Azure AI Speech LLM with 25 additional languages, up to 2x faster processing, and improved transcription accuracy.

Anthropic launched a dedicated plugin for verified nonprofits featuring specialized workflow skills for fundraising, grant research, marketing, and hiring.

OpenAI launched GPT-6.1 Terra, combining visual reasoning and computer use with 50% lower standard API prices.

Anthropic rolled out the unified chat and Cowork experience to Claude Enterprise alongside role-based access controls and an Efficient mode.

Moonshot AI launched a high-speed variant of Kimi K3.1 Pro offering the same intelligence at approximately two to three times the speed.

MiniMax published open weights for M3.1 Flash, a 428B multimodal mixture-of-experts model with 23B active parameters and a 1M context window.

An open small language model optimized for conversational agents, fitting in about 4GB of memory with 64K context support.

Moonshot AI released Kimi K3.1 Pro, an updated multimodal reasoning model with a 1M context window and improved coding capabilities.

Canva's first in-house image foundation model, generating high-quality 2K resolution images across any aspect ratio via Canva and public API.

A new API endpoint that returns quick typed answers, like a yes-or-no probability or a pick from fixed options, instead of free text.

Cohere launched Rerank 5 in Pro and Fast tiers, featuring Context Focus passage extraction and an expanded 32k context window.

Anthropic launched Claude Duet in research preview, an app where Claude operates as a proactive personal agent across connected tools.

Recursive introduced Flint, an open-weight 40B mixture-of-experts model activating 2.5B parameters per token for fast local coding agents.

Mistral AI released Timestral, an open-weight foundation model family delivering zero-shot probabilistic forecasting for multivariate time series.

Google released Gemini 4 Argon, a frontier model delivering advanced performance in software engineering, cybersecurity, and enterprise knowledge work.

Anthropic calls it the cheapest, fastest, and most capable small model it has released, at a much lower price than Haiku 4.5.

Mistral AI launched a preview of Mistral Large 4, a 1-trillion parameter multimodal model with 52 billion active parameters.

OpenAI introduced Dots, persistent always-on AI agents in ChatGPT powered by GPT-6 Astra for long-running work.

OpenAI launched GPT-6.1 Sol, bringing near-Astra intelligence for coding and computer use at one-fifth the token price.

Anthropic released Claude Sonnet 5.5, which runs over 30% faster and costs up to 30% less than Claude Sonnet 5.

The first Claude 5.5 model performs at Fable 5.1 level on most work and costs 40% less to run than Opus 5.

OpenAI launched GPT-6 Sol and Luna, bringing frontier intelligence to everyday work with lower latency and reduced API pricing.

Google introduced Gemini 3.8 Live and Extended Thinking to enable near real-time reasoning for voice agents.

OpenAI released GPT-6 Astra, featuring frontier reasoning, coding, cybersecurity, and computer use capabilities.