Deep Currents 04.04.26

Reading the Currents

The goal of these posts is to help you keep your head above water. This first section, called Reading the Currents, covers what I think matters most from the past month’s developments. If you want the full rundown of updates, slide down to The Full Stream.

The agentic moment

Something shifted in AI over the past two months, and in March the consequences became clear. To understand why every major lab simultaneously shipped agentic features, killed products, declared emergencies, and revealed plans in accidentally leaked roadmaps, you have to start with a claw.

OpenClaw, the open-source AI agent framework created by Austrian developer Peter Steinberger, went from a side project to a global phenomenon in a matter of weeks. Originally called Clawdbot (until Anthropic sent a trademark complaint), then briefly Moltbot, OpenClaw surpassed 250,000 GitHub stars faster than any open-source project in history, blowing past React and even Linux. In China, people lined up outside Tencent and Baidu headquarters to get help installing it on their laptops. OpenRouter reported that token consumption more than doubled in a single month, directly tied to OpenClaw’s growth. Nvidia compared its significance to what GPT was for chatbots.

OpenClaw’s appeal is simple but compelling: it doesn’t just chat, it can do things. It connects to your messaging apps, your calendar, your email, your file system, and it can execute tasks on your machine, and on the internet. It runs locally, works with any model, and has a growing marketplace of community-built skills. It is, in short, the proof of concept that many people have been waiting for: an AI agent that runs 24 hours a day and operates across everything you use.

OpenAI moved first on the talent. In February, Sam Altman announced that Peter Steinberger was joining OpenAI to “drive the next generation of personal agents.” OpenClaw would remain open-source in a foundation, but the person who built it would now begin working on what comes next.

Then another wave caught OpenAI off-guard: Anthropic stood up against the Pentagon during recent contract negotiations and in doing so positioned themselves as the ethical AI company, while Sam Altman’s opportunistic and rushed contract-signing with the Department of War cast OpenAI as the bad guys, and suddenly a QuitGPT movement was born. Combine that with Claude Code becoming the most popular coding app among developers, and Claude Cowork offering business users a similarly powerful agent experience in a user-friendly desktop interface, and Claude had suddenly become a serious threat to OpenAI’s dominance in the enterprise market.

At a recent all-hands meeting, OpenAI’s Chief of Applications Fidji Simo described the company as operating in a “Code Red” state, warning staff against “side quests” and citing Anthropic’s gains as the reason for sharper focus. The data is telling: Anthropic now captures more than 73% of all spending among companies buying AI tools for the first time, and for several weeks Claude overtook ChatGPT as the most downloaded app in the US. OpenAI’s response, still under development, is being billed as a “super app” merging ChatGPT, Codex, and the Atlas browser into a single unified desktop application.

And they killed Sora. The standalone AI video app was burning roughly $15 million per day in compute costs while user numbers collapsed. Disney, which had committed to a $1 billion investment and character licensing deal in December, learned about the shutdown less than an hour before the public. No money ever changed hands. But the real reason was because every GPU cycle spent rendering consumer videos was a cycle not spent on the agentic tools that actually drive revenue. With both OpenAI and Anthropic rumoured to be preparing IPOs before year-end, the race to prove enterprise dominance is critical.

OpenAI expanded Codex with security scanning and plugins for Slack, Figma, Notion, and 20+ other tools. Meanwhile, Anthropic kept shipping. Claude Code and Cowork got computer use, scheduled tasks, and Dispatch for controlling your desktop from your phone. New features were being released almost daily.

But it wasn't just the two frontrunners. The agentic wave hit everyone at once: Cursor released agent automations and a plugin marketplace connecting to 30 external systems. Lovable added several new tools that go beyond building apps. Even Slack turned its bot into an MCP-connected agent platform. All within the same month.

The competitive axis has shifted. It used to be about which model was smartest. Now it’s about which agent can reach the most things. Computer use, plugin ecosystems, background execution, cross-device control... These are platform plays. The coding agent that can also review your PRs, control your browser, respond to your Slack messages, and keep working while you sleep is a fundamentally different proposition than a chatbot that writes good code.

What the Anthropic leaks reveal

Into this frantic and chaotic competitive landscape, Anthropic accidentally allowed two data leaks within five days. On March 27, security researchers discovered nearly 3,000 unpublished assets sitting in a publicly searchable data store, exposed by a CMS misconfiguration. Among them: a draft blog post describing a new model called Claude Mythos. Then, on March 31, a routine Claude Code update shipped with a source map file pointing to a zip archive containing the tool’s entire source code. Around 500,000 lines across roughly 1,900 files, quickly mirrored across GitHub and dissected by the developer community.

Neither leak exposed customer data. Both were attributed to human error. But the timing was awkward for a company that prides itself on being responsible, and whose upcoming model is being positioned specifically around cybersecurity.

What makes these leaks particularly interesting is what they reveal about where all of this is headed.

The Mythos draft described a new model tier called Capybara, sitting above Opus as Anthropic’s most powerful offering. According to the leaked materials, it represents a significant jump in reasoning, coding, and cybersecurity capabilities. Anthropic has confirmed the model exists and is being tested with early-access customers, calling it “a step change” in performance. The company has also been privately warning government officials that Mythos makes large-scale cyberattacks significantly more likely in 2026.

The Claude Code leak was even more revealing. Beyond confirming the Capybara/Mythos model is actively being prepared for launch, the source code exposed dozens of unshipped feature flags, including one called KAIROS (referenced over 150 times), which appears to be an autonomous daemon mode that lets Claude Code operate as an always-on background agent, performing “memory consolidation” while the user is idle. Other unreleased features include the ability for Claude to review its own sessions and transfer learnings across conversations, and a persistent assistant that keeps working when the user steps away.

Connecting the dots

OpenClaw proved that people want agents that persist, learn, and operate across their entire digital life. Every major lab is now racing to build exactly that. And Anthropic’s leaked roadmap shows they’re already building features that go well beyond what any of them have shipped publicly: Agents that learn, that remember, and that work while you sleep.

The fact that this vision was confirmed in accidental npm packaging errors and misconfigured CMS settings reveals just how fast and loose the big labs are moving right now, and how carefully people are watching them.

And the OpenClaw story took one more turn as this newsletter was going out. Late on Friday, Anthropic announced that starting today, Claude Pro and Max subscribers can no longer use their subscriptions to power third-party tools like OpenClaw. Users will need to switch to Extended Use pay-as-you-go billing or use a separate API key. Anthropic explained it was due to the strain that always-on agent traffic was placing on their infrastructure, noting that subscriptions weren't built for those usage patterns. Steinberger, now at OpenAI, pushed back publicly, and OpenAI's Codex team was quick to signal that they'd welcome the displaced users. It's a move that makes economic sense for Anthropic, but it carries real risk: the open-source community that helped fuel Claude's popularity is now being told to pay up or move on. Many might decide to do the latter when faced with the true cost of running a 24/7 agent.

The design tools (almost) no one is talking about

One more current worth mentioning, even though it’s quieter than the agent wars: A steady stream of updates is making creative tools programmable in ways that impact anyone who works in design, media, or content production.

Figma updated its MCP integration so that agents can now write to the canvas, not just read from it. ElevenLabs launched ElevenCreative with Flows, a node-based pipeline builder for chaining creative models together. Google revamped Stitch as a vibe-design platform. Paper shipped MCP server support. Pencil continues to iterate with a steady stream of enhancements and improvements. And investors are clearly paying attention: a new startup called Noon just raised $44M for a canvas-based design tool that works directly on production code.

These aren’t big splashy announcements, and most people outside design and media likely won’t care, but they’re part of the same story. The agent revolution isn’t limited to coding and enterprise workflows. It’s reaching into every professional discipline where production work has traditionally been manual, specialized, time-consuming, and expensive. The design canvas is becoming programmable, and the pipeline from ideas to finished assets is getting shorter with every update. For now at least, humans are directing the agents, and they're still the final deciders. It will be interesting to see if, or how long, that lasts.

The Full Stream

Here’s the full rundown of the past month's releases across every category.

Agents & Agentic Coding

Alibaba released Qwen3.6-Plus which was optimized for “vibe coding” with greatly improved agentic coding capabilities.

Arcee released Trinity-Large-Thinking, a frontier open reasoning model for complex, long-horizon agents and multi-turn tool calling. It scored #2 on PinchBench, a benchmark from Kilo measuring model capability on tasks relevant to agents like OpenClaw, just behind Opus-4.6 but ~96% cheaper.

Anthropic shipped a wave of new Claude Code features: scheduled tasks on the desktop app (running as long as the computer is awake) paired with a loop command for recurring prompts over up to 3 days, Code Review which dispatches a team of agents on every PR to catch bugs, Auto mode as a safer alternative to the “–dangerously-skip-permissions” flag. Computer Use was released for both Cowork and Claude Code (CLI and desktop). Both tools also got Dispatch, a remote use feature connecting to the desktop app via the Claude mobile app, and Claude Code got Channels for pushing messages, alerts, and webhooks from Telegram, Discord, and iMessage into your session.

Cursor launched Cursor 3, rebuilt from the ground up as an agent-based development platform. They also released agent automations and announced 30 plugins and a plugin marketplace for connecting agents directly to other systems, and an updated Composer 2 coding model that competes with Opus 4.6 and GPT-Codex-5.4.

Google revamped Google AI Studio as a vibe-coding platform using Antigravity and interactive Firebase.

H Company released Holo3, a computer-use agent with just 10B active parameters that set a new record on the OSWorld-Verified benchmark, the leading desktop automation benchmark test.

Lovable added several new tools that go beyond building apps to let you analyze data, create and edit files and docs, and turn things like spreadsheets into full-stack apps. They also partnered with Aikido to provide penetration testing.

OpenAI released Codex Security, an agentic tool that scans repositories, builds threat models, and proposes context-aware patches for vulnerabilities. They also added plugins to Codex for Slack, Figma, Notion, Gmail, Google Drive, and 20+ other tools, meaning Codex can now handle planning, research, and coordination across work apps. These plugins are available in the Codex app, CLI, and VS Code extension.

Simular launched Sai, a new always-on agentic co-worker.

Slack released a major update to its Slackbot, turning it into an agent-powered platform for centralized knowledge that can connect to other platforms via MCP, gaining 30+ new capabilities, including meeting transcription, reusable AI skills, and a native CRM.

Design & Creative Suites

ElevenLabs launched Flows, a node-based canvas inside ElevenCreative for exploring creative asset generation by chaining models together and batch executing entire creative pipelines. They also launched a teams option with shared workspaces, shared credit pools, a centralized content library, and more affordable Basic Seats.

Figma updated the official Figma MCP to allow agents like Claude Code to directly manipulate elements on the canvas. Previously they could only read and retrieve data.

Google revamped Stitch as a vibe-design platform.

Noon raised $44M for their canvas-based design tool that works directly on production code. Currently accepting signups for early access.

Paper launched a desktop app with MCP server support along with many other enhancements and fixes.

Pencil added mesh gradients, absolute positions in flex layout, slides and presentation mode, and numerous other enhancements including updates to the Pencil documentation site.

Frontier Models

Alibaba released Qwen3.5-Omni, a new multimodal AI that processes text, images, audio, and video. Natively pretrained on massive amounts of text, visual data, and more than 100 million hours of audio-visual content, it supports speech recognition in 113 languages and dialects and speech generation in 36.

Anthropic updated Claude’s context window to 1M tokens for Opus 4.6 and Sonnet 4.6 by default. They also updated Claude's chat tool to be able to generate interactive charts, diagrams, and visualizations directly within the conversation (vs a standalone Artifact). They also added workplace tool connectors for Figma, Canva, and more to the Claude mobile app.

Google released Gemma 4 in four sizes, including a 31B model that ranked #3 among all open models and an edge model that runs on a Raspberry Pi with only 1.5GB of memory.

MiniMax launched M2.7, what the company calls its “first model which deeply participated in its own evolution,” writing its own training code, running autonomous improvement loops, and matching the scores of top Western models.

OpenAI released GPT-5.4 with native computer use capabilities, and GPT-5.4 Mini and Nano variants.

Images

Flora added an image editor to the generation workflow, allowing composite, inpaint, outpaint, resize, and split into layers.

Luma released Uni-1, a multimodal reasoning model that generates bitmap images and reached #1 on the Elo human-preference leaderboards for overall quality, style, and reference-based generation. It can also edit and combine up to nine reference images.

Midjourney released V8 alpha claiming better direction-following, much improved text rendering, and image rendering up to 5x faster than V7. The next release is likely to be named 8.1 instead of “V8 beta.”

Mergers & Acquisitions

Dreamer announced it is licensing its tech to Meta, with its full team joining Meta Superintelligence Labs in an undisclosed deal.

Microsoft acquired Cove, which had developed a collaborative workspace for creating AI apps.

Music

Google launched Lyria Pro 3 which can generate complete songs up to 3 minutes long.

Suno released version 5.5 with Voices and two new personalization features. Custom Models lets you create a fine-tuned model based on at least six songs from your Suno library, and My Taste adds a magic wand tool to prepopulate the Styles field based on your most used tags.

Productivity

Google upgraded Gemini’s capabilities in Workspace for Docs, Sheets, Slides, and Drive.

Littlebird is an always-on AI screen recorder that captures the text in your active windows and optionally the audio during your meetings, for persistent memory and context. Everything is encrypted and stored in the cloud on AWS servers.

Research Tools

Microsoft launched Critique and Council, two new features for Copilot’s deep research tool that leverage frontier models from Anthropic and OpenAI to assess and improve final output quality and accuracy.

Studies

Anthropic released the outcome of a massive qualitative study, leveraging their Interviewer tool to survey over 80,000 users about their usage and feelings about AI. It revealed a growing divide between basic chat users and AI power-users.

Video

CapCut released CapCut Video Studio with Dreamina Seedance 2.0.

Google released Veo 3.1 Lite, a new budget-friendly video generation model at less than half the cost of Veo 3.1 Fast.

Lightricks released LTX-2.3, an upgrade to its open-source video model, and LTX Desktop, a free local video editor built on the same engine.

OpenAI announced the shutdown of Sora, its standalone AI video generation app, just six months after launch. The app closes April 26, with the API following on September 24.

Voice & Transcription

Cohere released Transcribe, an open-source speech-to-text model that ranked #1 on HuggingFace’s leaderboard, with 14-language support and on-premise deployment.

Fish Audio released S2, which generates expressive speech with sub-150ms latency, multi-speaker in one pass, and inline emotion tags like [laugh] and [whisper] for rapid voice cloning.

Hume AI open-sourced TADA, a speech generation model that keeps text and audio in sync to reduce hallucinations while being 5x faster than similar models and small enough for on-device use.

Microsoft released MUI-Transcript-1 which they claim to be the most accurate transcription model in the world across 25 languages, excelling in noisy environments, and the best price-to-performance ratio of any cloud provider.

Mistral updated Voxtral TTS which can now clone a voice from a 5-second sample and generate speech in 9 languages on about 3GB of RAM.

That’s it for now! As always, please reach out if you have questions or thoughts to share, and let me know if there's something I missed.

This post was written with assistance from Claude as a research and writing collaborator. The hero image was generated with Midjourney.