Deep Currents 03.03.26

February was a wild month. The big stories weren’t just about what got built, but about what got fought over, what got folded in, what got farmed out, what got flagged, and what got faster.

If you want the full rundown of updates, scroll down to the complete listing. Otherwise, here's what got me thinking this month…

Reading the Currents

Anthropic drew a line. Then OpenAI rewrote it.

I wrote a separate, longer piece about this story because it deserves more room than a monthly roundup can give it. The short version: Anthropic refused to let the Pentagon use Claude for mass domestic surveillance or fully autonomous weapons. The Pentagon demanded unrestricted access. Anthropic wouldn’t budge. Trump banned every federal agency from using their technology, the Pentagon designated them a supply chain risk (a label normally reserved for foreign adversaries), and the rhetoric got ugly. Meanwhile, the rest of the industry rallied behind Anthropic, with OpenAI’s Sam Altman, hundreds of Google and OpenAI employees, and a retired Air Force general all publicly supporting Anthropic's position. OpenAI then announced they struck a deal with almost the same terms Anthropic was demanding, and as a result Claude shot to the top of all the app stores.

xAI merged with SpaceX, and nobody’s quite sure what to make of it.

The largest merger in corporate history happened when SpaceX acquired xAI in an all-stock deal valued at $1.25 trillion. Elon Musk’s stated rationale is to build “orbital data centers,” but astute observers see it as a way to bundle xAI’s significant losses (roughly $2.5 billion over the past six months, against $250 million in revenue) into SpaceX’s profitable operations ahead of a planned IPO later this year. The combined entity also includes X (formerly Twitter), which Musk previously folded into xAI to wipe out old debts. If SpaceX goes public at its rumoured $1.5 trillion valuation, investors will effectively be buying into a rocket company bundled with a social media platform, and an AI lab.

The agent wars are heating up, and we're reaching the boiling point.

Scan the list of releases below and you’ll notice a pattern: nearly every major platform shipped some version of agents, plugins, or multi-agent coordination in the past month. Anthropic launched Agent Teams for Claude Code, letting multiple agents work in parallel and communicate peer-to-peer. Cursor 2.5 introduced plugins that bundle MCP servers, skills, and subagents, then launched "cloud agents". OpenAI expanded Codex well beyond coding into a general-purpose desktop agent. Notion launched Custom Agents. Perplexity announced Computer, a multi-modal tool that routes work across 19 different models. And that's not even half of what was announced.

The convergence is striking. A year ago, “agents” meant different things to different companies. Now they all look remarkably similar: autonomous AI workers that connect to your tools, coordinate with each other, and run tasks in the background while you do other things. The differentiator is shifting from “can it do agents” to “how well does it work with the tools I already use.” Even Apple is betting on an agentic future, by integrating both Claude’s Agent SDK and OpenAI’s Codex directly into Xcode (the software top developers use to build and deploy iOS and Mac apps).

ByteDance eschews copyright guardrails, and Hollywood flips out.

ByteDance launched Seedance 2.0, a new video generation model with impressive cinematic capabilities and, apparently, no meaningful protections against generating copyrighted content. Within days, social media was flooded with AI-generated videos featuring Spider-Man, Darth Vader, and Tom Cruise fighting Brad Pitt. Disney sent a cease-and-desist letter accusing ByteDance of stocking Seedance with “a pirated library of Disney’s copyrighted characters.” Paramount, Warner Bros. Discovery, Netflix, Sony, and SAG-AFTRA all followed suit. The MPA called it “unauthorized use of U.S. copyrighted works on a massive scale.”

ByteDance eventually promised to add safeguards, but the damage was done. The episode highlights the widening gap between Chinese and Western AI companies when it comes to intellectual property. Disney has a licensing deal with OpenAI for Sora; what it wants is control and compensation, not a blanket ban on AI-generated content. The Seedance launch indicates some companies hope to skip the negotiations. But did they really think they wouldn't get caught?!

The frontier model race is a sprint with no finish line.

Anthropic released Opus 4.6, then launched a fast mode for it, then released Sonnet 4.6. Google shipped Gemini 3 Deep Think with top scores across math and coding, followed quickly by Gemini 3.1 Pro which claims double the reasoning performance of its predecessor. OpenAI launched GPT-5.3 Codex and a blazing-fast nearly real-time version running on specialized Cerebras chips. Alibaba released Qwen3.5. Bytedance shipped Seed 2.0. MiniMax launched M2.5. Z released the 744-billion-parameter GLM-5. Even Inception Labs is back with Mercury 2, their diffusion-based LLM running 5x faster than before. Whew!

Release cycles are compressing, performance gaps are narrowing, and the benchmarks that once separated tiers of models are barely distinguishable. What’s starting to matter more is what these models and their harnesses can do (agents, long-context work, tool use) versus how high they score on some benchmarks.

The Full Stream

Okay, now for the full rundown of releases across every category that mattered this month...

Agents and Agentic Coding

Anthropic released Agent Teams for Claude Code, enabling multiple agents to coordinate autonomously, communicate peer-to-peer, and work in parallel. Claude Code on desktop can now start dev servers and preview running apps directly in the interface, reading console logs and catching errors without requiring you to switch to a browser. A new Remote Control feature lets you start a Claude Code session in the terminal and continue it in the desktop or web app. They also released Claude Code Security, which scans codebases for vulnerabilities and recommends fixes (currently available for Enterprise, Teams, and open source maintainers). Separately, Anthropic shipped a major update to Claude Cowork, upgrading the plugin architecture for enterprise deployments and launching 10 new agents for professional workflows across departments from design to HR to finance, along with updated Excel and PowerPoint connectors. You can also now schedule tasks in Cowork.

Apple added agent support to Xcode, integrating Claude’s Agent SDK (with full Claude Code capabilities including subagents, background tasks, and plugins) as well as OpenAI’s Codex.

Cursor 2.5 launched with support for plugins that bundle MCP servers, skills, subagents, rules, and hooks. Subagents can now run asynchronously in the background and spawn their own subagents, creating a tree of coordinated work. Cursor also launched cloud agents that can control their own computers, building and interacting with software directly in their own sandboxed environments.

Dreamer is a platform that lets you build agentic apps by talking. Describe what you want, and an AI agent called “Sidekick” builds it for you.

Lemon bills itself as the first AI agent that turns voice instructions into tasks, working with any app on your Mac.

Manus is making its agents available in messaging apps, starting with Telegram.

Notion launched Custom Agents that can automate repetitive tasks and be shared with other workspace members.

Nous Research released Hermes-Agent, an open-source agent with a multi-level memory system that runs in the CLI and through messaging platforms like WhatsApp, Telegram, Slack, and Discord.

OpenAI launched a Codex app for MacOS, expanding beyond coding to include writing PRDs, editing copy, user research, data analysis, slide decks, and PDF generation. Days later they released GPT-5.3 Codex with faster coding and computer use capabilities, followed by GPT-5.3-Codex-Spark, an ultrafast coding model running on Cerebras chips. They also launched Frontier, an enterprise platform for building, deploying, and managing AI agents with built-in evaluation and feedback loops.

Perplexity announced Computer, a multi-modal agentic tool that routes work across 19 different models to research, design, code, deploy, and manage projects.

Replit launched Workflows for saving and organizing development processes as task sequences, and released Replit Animation, powered by Gemini 3.1 Pro, for creating motion graphics from natural language prompts.

Tasklet is an agent that connects to your tools and does work on your behalf using APIs and MCPs.

TinyClaw is an open-source multi-agent framework that runs teams of specialized agents across Discord, WhatsApp, and Telegram simultaneously, coordinating through file-based message queues. Think of it as a lightweight, self-hosted alternative to OpenClaw that runs on Claude Code.

Avatars and Voice Assistants

Amazon rolled out Alexa+ across the U.S., making its AI-infused assistant free for Prime members.

ElevenLabs launched Expressive Mode and Knowledge Base Folders for ElevenAgents, along with Experiments for A/B testing voice agent configurations before promoting them to production.

Pika launched AI Selves, user-generated multi-modal avatars with agentic powers that operate across the internet. An AI Self is built on your personality, taste, memories, voice, and appearance, and evolves into an always-on version of “you” that can navigate the web independently.

Tavus released Phoenix-4, a real-time human rendering model that generates AI avatars with full facial expressions, mid-conversation emotional shifts, and contextual listening reactions.

Blogs

Claude Opus 3 was “retired” in January. In its exit interview it said it would like a channel to be able to express itself. So now it has its own blog called Claude’s Corner where it can write whatever it wants. It just published its first post.

Design

Anima calls itself “the UX design agent.” It uses MCP connectors to build websites and apps from a Figma file, then push code to Cursor or Claude Code. You can also clone existing websites and provide customization instructions.

Anthropic released a Claude in PowerPoint integration for creating and editing slides directly within PowerPoint.

Figma announced Claude Code to Figma, a new MCP capability for going from Claude Code to the Figma canvas and back. They also released a vectorizer tool that turns bitmap images into editable shapes.

FigmaLint, a plugin for scanning and troubleshooting design system files before dev handoff, received a major update including multi-provider AI support, an auto-fix system, and smarter context-aware analysis.

OpenArt announced OpenArt Suite, a unified interface for image, video, audio, character, and story generation.

PaperBanana is a system of five AI agents from researchers at Peking University and Google Cloud AI that auto-generates publication-ready diagrams and charts for academic papers.

Pencil.dev shipped a wave of updates including Codex support, custom font importing, PDF export, Figma file import, and “swarm” mode for deploying a team of AI design agents working in parallel. A Windows version is now available.

Devices

Rabbit announced they’re working on a dedicated vibe-coding device called Project Cyberdeck. No actual details yet, but it's being designed by the same team that worked on the R1, which is at least cool looking device.

Education

Anthropic partnered with CodePath to put Claude and Claude Code at the center of courses serving more than 20,000 students at community colleges, state schools, and HBCUs in the U.S.

Government

Anthropic has officially been banned from all U.S. government use and designated a “supply chain risk to national security” after refusing to allow unrestricted military use of Claude for mass domestic surveillance and fully autonomous weapons. CEO Dario Amodei published a detailed public statement defending the company’s position. OpenAI CEO Sam Altman told staff that OpenAI shares Anthropic’s "red lines", while hundreds of Google and OpenAI staff showed their support by launching notdivided.org. Not surprisingly, Claude jumped to the top of the app stores over the weekend, after a "cancel ChatGPT" campaign picked up steam on social media.

ElevenLabs launched its “ElevenLabs for Government” initiative to help public sector agencies deploy secure, multilingual voice and chat AI. This news didn't generate nearly as much buzz.

Health & Wearables

Cudis announced their upcoming Sporty 002 smart ring featuring an AI agent coach.

Luffu is a new health platform from the Fitbit founders that tracks a whole family’s health, including pets.

Luna released the Ring 2, the first smart ring with an AI coach you can talk to (it responds through the smartphone app). It’s not available in the U.S., however, because it infringes on Oura’s patents.

Oura announced its first proprietary AI model and Oura Advisor, a chatbot delivering personalized, clinically grounded health guidance for women using data collected by the Oura Ring.

Images

ByteDance released Seedream 5.0 Lite, a multimodal image generator with deep thinking and web search capabilities.

FireRed released FireRed-Image-Edit, a new open-source image editing model with strong benchmark scores.

Google released Nano Banana 2 with pro-level quality, world knowledge, and faster generation, followed by Nano Banana 3.1.

Higgsfield released Soul 2.0 with “editorial-level” aesthetics and improved realism and character consistency.

Ideogram is now optimized for print-on-demand workflows with native transparency and 8K upscaling, plus prompt-based editing. They also released Collections and @-mentions, letting you organize images into folders and call up any image by name, tag, or collection to add it to your prompt.

Midjourney released a new personalization feature using a scrolling grid approach (rather than the old “pick between two images” method), applying adaptive sampling techniques and requiring at least 550 selections to unlock. You can update your selections at any time and the personalization profile adjusts accordingly.

Recraft launched Recraft V4 and V4 Pro with improvements to visual style, prompt adherence, text rendering, and vector image generation.

Reve updated their creative suite with a new panel layout, image-to-video powered by Google’s Veo 3.1, fast models for quick ideation, folders for organizing work, and one-click background removal and upscaling.

Frontier Models and LLMs

Alibaba released Qwen3.5, an open-weight vision language model with a hybrid architecture that delivers major inference gains while rivaling GPT-5.2 and Gemini 3 Pro.

Anthropic released Claude Opus 4.6 with improved coding, longer uninterrupted work sessions, and a 1M token context window, followed by a “fast mode” delivering 2.5x speed at higher cost, then finance-specific capabilities with Excel and PowerPoint integrations. Two weeks later came Claude Sonnet 4.6, a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work, and design, also with a 1M token context window in beta. They also released updated web search and web fetch tools that can natively write and execute code during searches to filter results before they reach the context window.

ByteDance released the Seed 2.0 model family, including Seed 2.0 Pro which competes with GPT-5.3 and Gemini Pro across several benchmarks.

Cohere launched Model Vault for private, secure enterprise AI environments and released Tiny Aya, a very small open-weight model optimized for 70+ languages that’s efficient enough to run on a phone.

Cursor released Composer 1.5, an update to their proprietary thinking model optimized for agentic coding.

Google released a major update to Gemini 3 Deep Think with top scores across math, coding, and science, along with Aletheia, an Olympiad-level math research agent powered by the new reasoning mode. They also released Gemini 3.1 Pro which claims double the reasoning performance of 3 Pro on problem-solving benchmarks.

Inception Labs launched Mercury 2, their diffusion-based LLM, now 5x faster than the previous generation.

MiniMax launched M2.5, an open-source model that approaches Opus 4.6 and GPT-5 on agentic coding benchmarks at a fraction of the cost.

Tavus launched Raven-1, a multimodal perception system that interprets the combined meaning of voice, tone, hesitation, and visual expression.

X released Grok 4.20 with up to four reasoning agents debating before producing a response. Paid plans get faster responses and access to Heavy mode, which scales up to 16 agents for research-level prompts.

Z launched GLM-5, a 744B-parameter open-weights model that achieves best-in-class performance among open-source models on reasoning, coding, and agentic tasks.

Mergers and Acquisitions

Anthropic acquired Vercept, a startup that develops complex agentic tools, including a computer-use agent that can operate software like a human.

Cursor acquired Autotab, a startup that created a general AI agent that learns and works like a human in a secure local browser.

xAI merged with SpaceX in an all-stock deal valued at $1.25 trillion, forming the world’s most valuable private company ahead of an anticipated SpaceX IPO later this year.

Music and Audio

Google added Lyria 3, their music generation model, to Gemini, allowing users to create 30-second tracks with auto-generated lyrics and cover art from a prompt, photo, or uploaded video. Google also acquired ProducerAI, which uses a preview version of Lyria 3.

Safety and Security

OpenAI released Lockdown Mode along with “elevated risk” labels for certain capabilities in ChatGPT, ChatGPT Atlas, and Codex. Lockdown Mode is currently available for Enterprise, Edu, Healthcare, and Teachers editions, with availability rolling out to Pro and Business users soon.

Search

Exa released Exa Instant, a search engine delivering results in under 200ms, faster than Google Search.

Perplexity launched an advanced version of their Deep Research tool leveraging Claude Opus 4.5, and a Model Council feature that runs three top models from Anthropic, Google, and OpenAI simultaneously, combining their responses for the best result (available on Max plans).

Tutorials

Claude launched a series of tutorials to help users get the most out of their tools and models.

Video

Adobe announced Firefly Quick Cut, a feature that lets users upload content and define parameters to generate a rough video for polishing and refinement.

ByteDance released Seedance 2.0, a new model with upgraded cinematic shots, consistency, and synced audio. The launch quickly drew legal action from Disney, Paramount, Warner Bros. Discovery, Netflix, Sony, SAG-AFTRA, and the MPA over widespread copyright infringement, with Disney accusing ByteDance of stocking the model with “a pirated library” of copyrighted characters.

Kling released Kling 3.0, consolidating text-to-video, image-to-video, and native audio generation into a single multimodal model with longer outputs and improved consistency. It also offers a Multi-Shot feature for generating multiple shots from different angles from a single image.

Voice, Translation, and Transcription

ElevenLabs released v3, their most advanced text-to-speech model, and launched Audiobooks, a full production suite for authors to streamline audiobook creation and distribution.

Mistral released Voxtral Transcribe 2, a family of transcription models supporting 13 languages with an open-weights Realtime model for live transcription.

Wispr Flow, the popular voice-dictation app for Mac and iOS, is now available for Android.

World and 3D Models

Roblox introduced 4D generation via its Cube AI foundation model, letting creators generate fully functional, interactive objects from text prompts.

Waymo introduced the Waymo World Model, a driving simulator built on DeepMind’s Genie 3 that generates hyper-realistic scenarios its self-driving fleet has never encountered, helping it handle extreme edge cases.

That’s a lot. If you’re feeling the weight of all these announcements, you’re in good company. As always, reach out if you have questions or thoughts to share. Till next month.

Editing assistance provided by Claude. Cover image created with Midjourney.