Deep Currents 11.11.25

Welcome to the latest installment of Deep Currents, a monthly curated digest of breakthroughs, product updates, and significant stories that have surfaced in the rapidly-evolving world of generative AI. My goal with these posts is to help you keep your head above water.

November brought another deluge of releases across every category of generative AI. The pace remains relentless, with a wave of announcements in design, agentic coding, and music licensing deals. If the list below seems overwhelming, feel free to jump to the end where I attempt to synthesize what all these updates mean.

Okay, let's dive into this month's currents…

Design and Marketing

Marketing and design tools saw new AI agents and tools specifically tailored for B2B workflows and creative production.

• Adobe launched a series of new AI agents specifically for B2B marketing teams at their annual MAX event, including Audience, Journey, and Data Insights systems. The broader Creative Cloud suite received conversational editing assistants, a new Firefly Image Model, broader access to third-party models, and new video tools.
• Canva announced Creative Operating System with a number of updates including a Canva Design Model trained specifically for creating design assets. This new model generates editable designs rather than flat images, applicable across various formats like social media posts, presentations, and websites.
• Figma acquired AI media generation startup Weavy for an undisclosed amount and will rebrand it as Figma Weave. Weavy's web tools enable users to combine different AI models and then edit each layer separately.
• Freepik announced Spaces, an infinite canvas where users can combine image, video, audio, retouch, upscale, and editing tools, all connected through visual nodes that represent steps in the creative flow.
• Google Cloud and Figma announced an expanded collaboration that brings more of Google's generative AI technology into Figma's design and product development platform along with faster image generation.
• Google released Pomelli, a new Labs experiment that designs AI marketing campaigns and content based on a brand's website.
• Miro announced the AI Innovation Workspace which includes the new AI Canvas product for visual AI workflows and Sidekicks (both waitlist-only), which are like AI-agent recipes for common and repeatable tasks.

Images

Image generation tools reached new levels of workflow sophistication, with visual interfaces and ready-made workflows making advanced techniques more accessible.

• Higgsfield AI launched Instadump, a new AI tool that converts a single photo into a collection of 15 professional shots.
• Krea released Krea Nodes, an intuitive visual interface that lets artists string multiple features and actions together into advanced workflows similar to ComfyUI.
• Leonardo launched Blueprints, a new workflow feature with ready-made AI workflows organized by categories like Photography, Products & Marketing, and Video & Motion that combine models, prompts, and settings for easier high-quality results.
• Microsoft released MAI-Image-1, which generates photorealistic images with particularly strong lighting effects and realism. It's available in the Copilot app.
• Recraft launched a new chat mode in beta, providing a new way to brainstorm, generate and iterate images, and run some multi-step design workflows.
• Stability AI and Electronic Arts announced a strategic partnership to co-develop generative AI models, tools, and workflows.

Video and Animation

Video generation maintained its breakneck pace with streaming generation, multi-minute outputs, and increasingly sophisticated editing tools.

• Eleven Labs launched Image and Video generation in beta, providing access to several top models including Sora 2 Pro, Google Veo 3.1, Kling 2.5, and Flux 1 Kontext Pro.
• Google updated the Flow video editing app. "Ingredients to Video" lets you make a video based on three reference images, "Frames to Video" creates a video from starting and ending images, and "Scene Extension" lets you add additional video up to a minute in length. All of these features generate full audio, powered by their launch of Veo 3.1.
• HeyGen announced Motion Designer, the world's first prompt-based motion graphics generator that transforms ideas into dynamic visuals, charts, explainer graphics, title cards, or social hooks aligned with your brand.
• LTX-2 is a new open-source model that generates both video and audio in one pass, supports up to 4K resolution at 50 fps, and offers two modes (Fast and Pro) depending on your use case.
• Meta added new AI tools to Instagram Stories, allowing users to restyle, edit, and remove objects from photos and videos directly within the platform.
• MiniMax's latest AI video model Hailuo 2.3 arrived with upgraded realism and motion.
• Odyssey launched Odyssey-2, a new interactive video model that generates streaming AI footage at 20 frames per second, allowing users to shape and control multi-minute videos through text prompts as they explore the scene.
• OpenAI updated Sora 2 with a new Storyboards feature that lets Pro users map out their video, added Character Cameos, and video stitching capabilities that connect multiple videos together to create longer, multi-scene clips. An Android app was also released.
• Runway launched a suite of task-specific tools called Apps, which are use-case specific features for creative editing like upscaling an image, reshooting products, or adding dialogue. They also announced Model Fine-tuning, giving users the ability to customize the company's generative video models on their own data and for specific use cases.

Audio, Music, & Voice

Audio tools continue to steadily improve while the music generation space saw big partnerships announcements as AI companies navigate licensing and rights challenges.

• Eleven Labs released ElevenLabs UI, a collection of open source agent and audio components that you can customize and extend. Their Voice Isolator now supports video, promising studio-quality audio no matter where you're recording by removing unwanted background noise.
• Inworld Labs released Inworld-TTS-1, a multilingual text-to-speech model that's fast and significantly cheaper than rivals.
• Hume launched a new Voice Conversion tool that lets you modify a pre-recorded voice and change it to a created or cloned voice while preserving intonation.
• Udio announced a partnership with Universal Music (UMG) with plans to launch a subscription service next year where fans can create remixes and customized tracks using licensed songs, where artists must opt in and generated songs stay within Udio's platform.
• Stability AI also announced a strategic alliance with UMG to develop next-generation professional music creation tools. Is Suno next?

3D and World Models

• ByteDance announced a new image-to-3D model for world building called Seed3D.

LLMs

Language models continued their march toward faster, cheaper, and more specialized capabilities, with several smaller models punching well above their weight.

• Alibaba's Qwen3-VL is a new "local" sized model that can run on a laptop, reading both images and videos to do things like extract info from receipts or turn screenshots into working code.
• Anthropic released Claude Haiku 4.5, their smallest and fastest model yet, which they claim is also their safest. They also expanded Claude's Memory feature beyond Team and Enterprise users to Pro and Max subscribers, so Claude can remember what matters across all your conversations without re-explaining your projects or preferences.
• Anthropic's Claude Desktop app gained several new capabilities including voice prompting, screen viewing on Mac, and connections to tools like code editors, local files, and databases to analyze data and automate tasks.
• Google released Gemini Deep Research which now connects to Gmail, Drive, Docs, and Chat to combine live web research with your internal documents in one report. They also upgraded NotebookLM with a larger context window, memory, customizable chat personas, and improved response quality. NotebookLM's Video Overview feature is now powered by Nano Banana for its presentation graphics, adding 6 new visual styles (like watercolor and anime) plus a new Brief format for quick insights.
• Google launched a File Search Tool for the Gemini API that provides a fully-managed RAG system for grounding chatbot responses in uploaded documents.
• Inception Labs launched a new version of Mercury, their Diffusion-based LLM, that provides higher quality than Claude 4.5 Haiku at one-fifth the latency and less than one-quarter the price.
• MiniMax released M2, a model optimized for agents and code that currently ranks as the top open weights model and #4 on Artificial Analysis' Intelligence Index.
• Moonshot AI released Kimi K2 Thinking, an open-source reasoning model that matches or exceeds models including GPT-5 and Claude 4.5 Sonnet across benchmarks at a much lower cost.
• OpenAI launched Company Knowledge, a feature for Business and Enterprise clients allowing users to access deeper context data living across company cloud platforms like Github, Hubspot, Slack, Sharepoint, and Google Drive. They also released an update to GPT-5 so it handles sensitive conversations better, recognizing and responding to users experiencing mental health emergencies, after consulting with over 170 mental health professionals in dozens of countries. And ChatGPT now lets you interrupt long-running deep research queries and add new context without restarting or losing progress by clicking "update" in the sidebar.
• Radical Numerics released RND1, an open-source 30B-parameter diffusion language model that generates entire sentences in one shot rather than one word at a time, outperforming other open diffusion models on math, coding, and reasoning tasks.
• Samsung released Tiny Recursion Model, a 7M-parameter open-source model that solves logic puzzles better than Gemini 2.5 Pro and o3-mini despite being 10,000x smaller.

Agents

The agent ecosystem expanded with pre-built skills, financial integrations, and tools for managing multiple agents across workflows.

• Anthropic launched Skills for Claude. Skills are folders that include instructions, scripts, and resources that Claude can load when needed. While working on tasks, Claude scans available Skills to find relevant matches and loads only the minimal information and files needed.
• Anthropic also expanded Claude for Financial Services with Claude for Excel (a beta tool that analyzes and modifies spreadsheets), new connectors to access real-time market data platforms like LSEG, Moody's, and Aiera, and six pre-built Agent Skills for tasks like building DCF models and initiating coverage reports.
• GitHub introduced Agent HQ, a platform that integrates coding agents from Anthropic, OpenAI, Google, Cognition, and xAI into existing workflows via a dashboard.
• Google released Gemini Enterprise, bundling its workplace AI offerings into a single platform where employees can create, deploy, and manage agents without coding experience.
• Amazon launched the Quick Suite AI platform, betting that companies want agents embedded directly in their workflows, not isolated in separate apps. It helps users automate tasks, do research, visualize data, and take action across enterprise apps.
• Manus released version 1.5 of its AI agent platform, with upgrades including 4x faster task completion, full-stack web development capabilities, and more.
• Microsoft released a new Workflows feature that automates routine tasks across the Copilot platform.
• Grammarly announced a complete rebrand of the parent company's identity to Superhuman, and launched a 'Superhuman Go' AI agent suite that connects with apps like Gmail, Jira, and Google Calendar to handle tasks like logging tickets and scheduling meetings.

Vibe Coding

Agentic coding tools continued their evolution with shareable plugins, cloud-based execution, and increasingly sophisticated deployment capabilities.

• Anthropic's Claude Code now supports shareable plugins, which will be the standard way to bundle and share Claude Code customizations. Claude Code is on the web now, allowing users to launch multiple Claude Code agents to do work in the cloud without needing to keep their machine awake.
• Factory AI lets you delegate entire coding tasks like refactoring legacy code, responding to incidents, or migrating databases to AI agents that work directly in your IDE, terminal, and Slack without changing your setup.
• CodeBanana aims to be the Google Docs but for code, allowing real-time collaboration with teammates, an AI that knows your code repo, and one-click publishing of apps.
• Cognition (maker of Windsurf) launched SWE-1.5, a new fast coding agent for exploring large codebases, building full-stack apps, and completing coding tasks 13x faster than other coding assistants. They also launched an AI code mapping tool called Codemaps.
• Cursor launched 2.0 and Composer, their first frontier-level agentic coding model, and a new interface for working with many agents in parallel. It now runs up to eight coding agents simultaneously on one prompt (each in isolated environments), plus adds voice control and team-wide custom commands managed from a central dashboard.
• Google expanded access to Opal to 15 more countries and now comes with debugging tools and speed improvements. They also added vibe coding to AI Studio, allowing users to turn simple text prompts into multi-modal AI apps without juggling multiple APIs or carrying out complex integration work.
• Lovable rolled out a new integration with Shopify, allowing users to build and launch online stores on the platform using natural language prompts.
• Microsoft launched App Builder that lets you create and edit full-stack applications using simple conversational prompts.
• Mistral launched Studio, a platform for companies to move from AI prototypes to production with built-in tools for performance tracking, testing, and security.
• OpenAI launched Aardvark, which scans your code repositories to find security vulnerabilities before attackers do, validates they're real by testing them, and generates one-click patches you can review and apply.
• Replit's new AI Integrations feature lets you build apps with AI features, offering 300+ models, including Gemini, OpenAI, Claude and open-weights models from OpenRouter, with the AI costs billed to your Replit account.
• Vercel's popular vibe coding tool v0 is now available on iOS.

Web Browsers

The browser wars gained a significant new entrant while existing players refined their AI capabilities.

• Dia launched a new Tabs tool to help users manage and interact with open browser tabs.
• OpenAI finally launched their rumored web browser, ChatGPT Atlas.
• Opera added a deep research agent to their Neon browser.
• Perplexity made several improvements to the Comet Assistant so it can work longer and on more complex jobs spanning larger periods of time.

Wearables

• Alibaba announced pricing for its upcoming artificial intelligence glasses, starting at $660 USD with pre-orders opening Oct. 24 and shipments beginning in December, though it seems to only be available in China for now.
• Sandbar, a new startup launched by ex-Meta engineers, announced preorders for a ring that lets you record your thoughts called Stream Ring, with an associated app for life-logging.

Education and Learning

• Google launched Skills, a learning platform featuring 3,000 AI and technical courses, with gamified features and employment pathways through company partnerships.
• xAI launched Grokipedia, an AI-driven Wikipedia-style encyclopedia with 800K+ Grok-generated articles and options to let users submit corrections with real-time AI edits.

What This All Means

November's developments reveal AI's continued march toward practical integration, but with some interesting inflection points that suggest the next phase won't just be "more of everything."

The music licensing deals between Udio, Stability AI, and Universal Music represent a significant shift from adversarial positioning to collaborative frameworks. These partnerships acknowledge that generative AI in music isn't going away, so the industry needs to build systems where artists (and their labels) maintain control and compensation. The approach of requiring artist opt-in and keeping generated content within platforms suggests a middle road between unrestricted generation and complete prohibition. Whether this model scales or collapses under real-world pressure remains to be seen.

The vibe coding space is reaching new levels of sophistication with tools like Cursor running up to eight parallel agents, and Claude Code moving to the cloud suggests we're approaching a phase transition in software development. When coding tools can work autonomously for hours, deploy backends automatically, and integrate directly with design systems, the bottleneck shifts from implementation to intention. The democratization of software development could unleash tremendous creativity, but it also raises concerns about code quality, security, and long-term maintainability when the people deploying applications don't fully understand how they work.

The launch of ChatGPT Atlas marks OpenAI's bet that the near-term future of web interaction relies on a familiar browser experience. Like other companies who have added AI sidebars to traditional browsers, OpenAI is building on existing conventions rather than redefining how we access and interact with web information. This somewhat mirrors their approach with Sora's social feed that imitates TikTok and focuses on mainstream consumption, versus traditional AI video platforms that cater entirely to hobbyists and professional creators.

The proliferation of tiny but capable models tells an interesting story about efficiency. When Samsung's 7M-parameter model can outperform Gemini 2.5 Pro on specific tasks despite being 10,000x smaller, it suggests we may have been over-parameterizing models for certain use cases. The push toward smaller, specialized models that can run locally on laptops represents a counter-argument to the "bigger is always better" mentality. This matters for privacy, cost, and accessibility.

The expansion of "skills" and pre-built agent workflows across multiple platforms (Anthropic's Skills, Leonardo's Blueprints, Miro's Sidekicks) reveals an industry recognizing that raw capability isn't enough. People need scaffolding, templates, and reusable patterns to actually benefit from AI tools. The shift from "here's a powerful model, figure it out" to "here are proven workflows you can customize" represents a maturation of the space.

Finally, the continued tension between open and closed approaches plays out across multiple domains this month. Open-source models like RND1, Kimi K2, and Samsung's Tiny Recursion Model compete directly with closed model offerings, while companies like Google and Anthropic simultaneously release both proprietary systems and open frameworks. This suggests the industry hasn't settled on a winning strategy yet, and we're likely to see continued experimentation with hybrid approaches that balance openness with commercial viability.

As has been the case all year, the pace remains exhausting, but November's releases suggest the industry is starting to grapple with questions beyond "what can we build?" and moving toward "how should this actually work in practice?" That should be seen as a sign of progress, even if it comes wrapped in an overwhelming flood of new releases.

Okay, that's enough for this month! As always, please reach out if you have questions or thoughts to share, or if you need any help making sense of all this.

Cover image created with Midjourney.