🚀 New AI releases

Mistral released the weights of its improved Mistral Large 4, a natively multimodal Mixture-of-Experts model with about 1T total and 49B active parameters per token, a 1M-token context, text and image input, and three reasoning modes. Mistral says it leads open-weight models in coding and defensive cybersecurity. Weights (FP8 and NVFP4) are on Hugging Face under the Mistral Ecosystem Open-Weight License, a modified MIT license that requires a separate commercial license for Model-as-a-Service businesses. Also on the Mistral API as mistral-large-4 at $0.70 input and $2.10 output per million tokens, and on OpenRouter · official source ↗

Canva's in-house vision-language model that looks at an existing design and makes the edits you ask for, from one line of copy to every slide in a deck, using an agentic loop that writes and runs code and then checks the page. Canva says it is up to 69% faster and 46% cheaper per edit than its previous design agent. Available now in Canva AI for Canva Pro subscribers (pick 'Brush' in the Canva AI dropdown), with wider availability coming · official source ↗

Pika's post-trained version of MiniMax H3 for fast, controllable video: up to 12 image, video and audio references in one generation, synchronized audio exported as up to five separate stems (dialogue, score, SFX, foley, ambience), a 768p draft mode with upscaling to 4K, and editing of characters, camera moves and voices. Pika says it can make a 15-second clip in about 7 seconds of compute. Available now through the Pika API, with open weights and inference code on GitHub and Hugging Face; coming to pika.art on November 2 · official source ↗

Copilot Fusion is now generally available: instead of picking a model, you choose Fusion at Standard, High or Max effort and Copilot routes each task to one model or orchestrates several (cascade or draft-critique-revise). GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index at about half the average API cost per task of Claude Code with Opus 5.5 Max. Billed per model used at standard rates (30% discount for a limited time on paid plans); available on all Copilot plans in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs · official source ↗

NVIDIA's open-weight transcription model (34B total, about 4.3B active parameters) for recorded audio up to one hour in 40 languages: verbatim or written-form transcripts steered by a plain-language instruction, speaker attribution, word timestamps, contextual transcription with expected names or terms, and speech translation. NVIDIA reports it is the best open-weights model on the Hugging Face Open ASR Leaderboard. Weights on Hugging Face (BF16, FP8, NVFP4) under the OpenMDW-1.1 license; works on recordings, not live streams · official source ↗

GitHub Agentic Workflows, an open-source GitHub Next project, is generally available: you write repository automations in plain-language Markdown, and they compile into GitHub Actions YAML for jobs like issue triage, CI failure analysis, docs updates and test writing, using your existing runners and policies. Available on GitHub Free, Team and Enterprise; the default Copilot CLI engine needs a Copilot license · official source ↗

Replit Cloud Agents let anyone on a team describe a change in plain language and get a tested pull request (GitHub) or merge request (GitLab) on the company's existing codebase. The agent tests the change in the running app in a cloud sandbox and attaches screenshots and video; engineers still review and approve. Shared environments are set up once by an admin, and requests can start from @Replit in Slack. Available now for all Replit Pro and Enterprise customers; skills-from-review-feedback is in early access · official source ↗

Microsoft AI's latest image editing model, launched in preview: restyle or relight whole scenes, change text and symbols inside an image, restore and enhance assets, and make edits that need an understanding of objects, lighting and environments. Microsoft says it improves on MAI Image 2.6 across all 10 use cases and 7 editing actions in Artificial Analysis's image-editing evaluation · official source ↗

Liquid AI's new generation of Liquid Foundation Models: six open-weight models of about 2B parameters on one shared LFM3-1.7B backbone, covering text, vision, audio, omni and two decision models. Built for on-device use (the text model fits in about 1 GB at 4-bit) with a 256K context window for text, and each can be fine-tuned on a single GPU. Available on Hugging Face, the Liquid Playground and the Liquid AI API, with llama.cpp, MLX and vLLM support · official source ↗

Higgsfield Playable turns a text prompt, a gameplay idea or your own assets into a playable 2D or 3D mobile-style game you can test in the browser, with 8 genres (shooter, runner, puzzle and more). Playable Games is on all Higgsfield plans and can also be generated through Claude with Higgsfield MCP; Enterprise customers can export a game as a playable ad (single HTML file) for 9 ad networks including AppLovin, Unity Ads, Meta and TikTok Pangle · official source ↗

OpenAI's new mid-tier GPT-6 model for everyday work: detailed visual reasoning, precise computer use and clearer, more natural writing. OpenAI says standard API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol, and it completes computer-use tasks over 3x faster than GPT-6 Astra and GPT-6.1 Sol; GPT-6.1 Sol and GPT-6 Astra remain the picks for the hardest coding. Available in ChatGPT Work, Codex and the API (gpt-6.1-terra), and on Amazon Bedrock · official source ↗

Describe a creative tool in plain language and App Studio builds a no-code app with its own interface that generates images, video and audio using your Runway Brand Kits and Workflows; existing Workflows can become an app in one click. Ships with built-in apps (Character Swap, AutoComp, Product Portfolio, 3D Camera Control, Precise Edit). Building is on all paid plans with no add-on, using existing credits; free users can run built-in and shared apps · official source ↗

Moonshot AI's new flagship, a post-trained version of Kimi K3 with a 1M-token context window, aimed at coding (including Unity, Unreal and Godot projects), 3D creation, financial research and agent tasks; the Highspeed variant runs 2-3x faster. Live on Kimi.com, the Kimi apps, Kimi Work, Kimi Code and the Kimi API; Moonshot says full open weights follow by 4 November and that cyber-related training data was removed for the open-weight release · official source ↗

Developers can now schedule agent runs in the OpenAI Agents API, either recurring with a cron expression and time zone or once at a set future time, starting a new session or continuing an existing one. Agents given the automations tool can also create, update or remove their own schedules from natural-language requests · official source ↗

Type @Gemini Notebook in a Gemini chat (or just ask) to create, edit or delete notebooks, add or remove sources, save notes and answer questions from your notebooks without leaving the chat. Available now to Google Workspace customers, Workspace Individual and personal Google accounts; on by default except in the EEA · official source ↗

Replit Agent can now act as an on-call engineer for published apps: with Monitoring alerts and Auto-fix on, it investigates production issues from the app's logs and metrics in the background and attempts and validates fixes, which you review and publish (or let an experimental Auto-republish setting ship). The same update lets Agent create reusable workspace skills and adds chat and asset sharing · official source ↗

Ideogram's first vector model generates clean, editable SVGs (logos, wordmarks, lettering, illustrations, labels) from a text prompt, or converts an existing image into a scalable vector, with editable colors, shapes and gradients for Illustrator, Figma or code. Live in the Ideogram app, the API (from US$0.10 per image conversion and US$0.13-0.32 per generated SVG) and partner platforms · official source ↗

OpenAI's API for fast bounded-answer decisions (classify, route, pick the next agent action) built on GPT-6 Luna is out of beta. New since beta: native GPT-Live integration for voice apps, prompt caching, native JSON inputs, images in question instructions, up to 400 questions per request and US/EU inference residency. US$0.10 per million input tokens (US$0.01 cached), no output-token charges · official source ↗

Canva Research's design-focused image model plans the layout, elements and text before rendering, at up to 2K resolution in 30+ aspect ratios, with text in 55 languages and outputs that convert into editable layers (Magic Layers). Available to all Canva users in Dream Lab and to developers through Canva's public API; Canva MCP support and faster variants are coming · official source ↗

Gemini-powered real-time speech translation in Google Meet now covers 63 languages on web, Android and iOS, auto-detects who is speaking which language and lets each participant turn translation on for themselves, with matching translated captions. Rolling out now to Rapid Release Workspace domains and from 2 November to Scheduled Release domains, on eligible Workspace and Google AI plans · official source ↗

Lets personal AI agents sign in to sites without ever seeing your password: the agent requests a login, you approve it on your device, and 1Password fills it in directly. Integrations with Meta Muse and Grok Bot rolling out now; Claude Duet, ChatGPT Work/dots and Town coming soon. No extra cost for 1Password Individual and Family accounts on 1Password.com · official source ↗

Open 14.8B model from Hugging Face and research partners that reads raw DNA inside a normal chat and answers in plain English: variant benign/pathogenic predictions, gene-regulation regions, organism classification and next-base prediction. Apache 2.0 weights, runs on a single 80 GB GPU with vLLM, ships with a Claude Code / Codex skill. Research use only, not clinically validated · official source ↗

Anthropic's new proactive agent app (iOS, Android and claude.ai), built around Claude Opus 5.5: a morning brief across your connected apps, tasks that keep running when you step away, and approval checks before it sends, pays, signs or deletes. Waitlist research preview for US Pro and Max subscribers, included in those plans' usage limits · official source ↗

Recursive's first model: a 40B-parameter mixture-of-experts coding model with about 2.5B active per token, built to run agentic coding on your own laptop (Recursive reports 125+ tokens/s on an M4 Pro MacBook Pro). Apache 2.0 weights on Hugging Face; works with llama.cpp, Ollama, LM Studio, vLLM and SGLang · official source ↗

New open, commercially usable retrieval models for agents and RAG: Nemotron 3.5 Embed (8B and 1B, text plus document images such as tables and charts), Rerank 8B, and a 631M-parameter Parse model that turns PDFs and scans into structured Markdown and tables. Weights on Hugging Face, plus NVIDIA NIM and partner APIs · official source ↗

Eleven Realtime v1 Mini (low-cost natural conversation) and v1 Max (reasoning and multi-step tool calls) are complete speech-to-speech voice agents in 90+ languages with any of 17,000+ library voices, available in ElevenAgents and through an OpenAI Realtime API-compatible WebSocket API · official source ↗

A new local-first mode in the Perplexity Mac app runs both the model and the agent on the Mac, working with local files and apps like Notes, Messages, Mail and Numbers, even offline; cloud access only when you approve. For Pro, Max and Enterprise subscribers, with Qwen 3.6 and PPLX Pro as local models; Apple silicon with 36GB+ unified memory recommended · official source ↗

New speech-to-text models (Voxtral Small and Mini Transcribe 3, plus a realtime model under 500 ms latency, 50+ languages with diarization) and Voxtral Speech 3 text-to-speech (35+ languages), plus a LiveKit-based voice agent SDK; Mistral API from US$0.0035/min for transcription and US$13 per 1M characters for speech, or self-deployed · official source ↗

Mosaic's video agent edits your footage into clips, makes launch videos and motion design, can copy a reference video's pacing and style, and exports to Premiere Pro, After Effects and DaVinci Resolve; also has an API and MCP server. Credit-based pricing · official source ↗

A new model behind Polly's Generative engine gives more natural, accurate low-latency speech (more reliable with numbers, emails and addresses), 14 new voices across nine locales, and Hindi and Finnish support · official source ↗

The Gemini app is starting to roll out a business landing page, saved brand styles for on-brand visuals with Nano Banana 2.1, financial charts from messy spreadsheets and receipts, and connections to Google Business Profile, Google Ads, Workspace and apps like Canva, Shopify, Wix and Xero, for all Gemini app users globally over the coming days · official source ↗

Notion's sixth agent harness gives Notion Agents their own computer and browser, better context management, a "Review for me" mode with safety checks and tasks of up to 24 hours. The upgraded Auto router picks a model per task (Notion reports Opus 5.5 / Astra-level results at about 90% lower cost), personal agents can use your ChatGPT plan, Notion MCP was overhauled, and Notion AI moves to a new credit system · official source ↗

Faster, cheaper agent and coding model: MiniMax says agentic task costs drop 70% versus its predecessor, with 200+ tokens/s output; US$0.15 input / US$0.60 output per 1M tokens via MiniMax API, M Plan and MiniMax Code (weights later) · official source ↗

Inception's largest diffusion LLM: 1,000+ tokens/s, 1M-token context, OpenAI-compatible API on Inception, OpenRouter, Baseten and Vercel; launch price US$0.10 input / US$0.35 output per 1M tokens · official source ↗

New Workers AI model that returns typed values (text spans, numbers, dates, estimates, forecasts) with confidence scores instead of generated text; US$0.10 per 1M input tokens, open weights planned · official source ↗

Gemini Notebook now suggests topics from your sources and lets you review the outline or producer notes before it renders a Slide Deck or Video Overview; rolling out to all users over the week · official source ↗

New V4.1 Flash checkpoint on web, app and API (same deepseek-flash name and price), faster and shorter outputs; DeepSeek's in-house search API opened to developers · official source ↗

Reflection's first open-weight model: sparse MoE (501B total, 23B active) for coding and agents, Apache 2.0 on Hugging Face, plus the open-source Mirror CLI · official source ↗

Create, co-edit, present and export decks (PowerPoint, PDF, Google Slides) in ChatGPT Space for Pro, Business and Enterprise; GPT-6 Pro gets faster replies and parallel agents · official source ↗

Describe a voice in text to create a custom voice for GPT-Live sessions (creation is free), plus nine new GPT-Live voices · official source ↗

Kling AI's new video model: native 30-second generation, steadier motion and camera moves, Omni Reference with up to 15 references and 10 keyframes; API coming soon · official source ↗

600B-parameter sparse MoE (27B active) with a 1M-token context, built for long agentic tasks; BF16 and FP8 weights on Hugging Face · official source ↗

Open-weight multimodal MoE model (816B total, 42B active, 1M context) for coding and agents · official source ↗

Open 120B multimodal agent model plus a 24GB local Bolt model · official source ↗

Automatic PR reviews out of preview for Team and Enterprise, about 3x faster · official source ↗

Upload a video and get a synced music score plus sound effects · official source ↗

Build apps with GPT-Live, Claude, Gemini 4 Argon, Perplexity, Firecrawl and more · official source ↗

Agentic review/verification of technical documents · official source ↗

xAI speech-to-speech model in the Voice Agent API, same US$0.08/min price · official source ↗

Google's frontier model opens to API, Pro/Ultra users · official source ↗

Single universal work agent inside Workspace · official source ↗

Interactive UI answers, rolling out to all tiers · official source ↗

Initial limited release to cyber defenders · official source ↗

OpenAI's new flagship model family · official source ↗