Deep Currents 05.05.26

Reading the Currents

This month brought another deluge of releases across every category. As always, the analysis comes first, followed by the full stream of everything else that was announced.

Memory Moves

Agents had a coming-of-age month, and they've started keeping track of things. Anthropic shipped Claude Managed Agents with built-in memory, stored as editable files you can review and modify directly. OpenAI rolled out a basic memory feature for Codex, and also released an experimental feature called Chronicle that captures your screen in the background to build persistent memory of what you've been doing. They also launched Workspace Agents, that can retain context across workflows in ChatGPT and Slack. Meanwhile Perplexity integrated its Computer agent with Plaid so it can see your bank accounts, credit cards, and loans to provide sophisticated financial analysis.

Closer to the Craft

Design tools had a similar moment. Adobe announced a Firefly AI Assistant that orchestrates multi-step workflows across Photoshop, Lightroom, Premiere, Express, and Firefly itself. Anthropic launched an impressive agentic prototyping and creative design tool called Claude Design, along with a broader suite of creative work connectors for 3rd-party tools including Affinity, Blender, Ableton, and Adobe's Creative Cloud. Figma released Weave, a node-based image and video generation workflow tool. And Google open-sourced the DESIGN.md spec so any coding agent can import or export a visual identity system.

The most intriguing update in this space however, came from Canva. CPO Cameron Adams said the new Canva Design Model was trained on "structured data, millions of designs, and the actual sequence of edits used to build them." That last part is key. While other models have been trained to learn what good designs look like, Canva's model is learning how designers get there. All the iterations, corrections, and small judgment calls that don't fit into a prompt just became training data. It's early days still but it will be interesting to see how this approach plays out.

The Floor and the Frontier

Open-weight model context windows hit a milestone this month. Alibaba's Qwen 3.6-Plus and DeepSeek's V4 both ship with 1M-token context. Eighteen months ago, a 1M context window was a frontier moat. Now it's the standard from Chinese open-weight labs.

The frontier closed labs didn't sit still mind you. Anthropic released Opus 4.7 and OpenAI shipped GPT-5.5 in the same window. Opus 4.7 came with mixed reviews and notably higher token costs for typical coding tasks. GPT-5.5 was, by most accounts, just better. The lesson, such as it is: the goalposts keep moving. Benchmarks and context length are catching up across the industry. The closed labs are betting that something else, call it judgment or taste or capability density, is what keeps them ahead. Additionally, the products they've built on top of these models, like Claude's Cowork, Code, and Design, and OpenAI's Codex, are able to leverage the capabilities of their latest models to deliver better outcomes, and ultimately provide the value that people will keeping paying for.

Each of these stories is a small advance. The accumulation is the story. Agents are getting more present in the work. The work is getting more present in the training data. The loop is tightening, and we're somewhere inside of it.

The Full Stream

Agents

The agent ecosystem moved from stateless assistants to systems that retain context across sessions, integrate with personal financial accounts, and operate inside team workflows.

• Anthropic opened a public beta for Claude Managed Agents, a platform that lets developers go from agent idea to live product. The same release added built-in memory, stored as editable files that can be reviewed or updated directly.
• Cursor released version 3.1, introducing a tiled Agents Window that runs multiple AI agents in draggable panes for side-by-side comparison.
• Google announced the Gemini Enterprise Agent Platform for governing thousands of agents across an organization.
• Microsoft added Copilot's agentic capabilities to Excel, PowerPoint, and Word, letting it perform multi-step actions directly inside documents.
• OpenAI gave their Mac desktop app a slate of new Codex capabilities including background computer use, image generation, 90+ connectors, a built-in web browser, automations, and memory. They also released Chronicle, a Codex preview feature that captures your screen in the background to build persistent memories, and Workspace Agents, shared bots designed to handle multi-step team workflows autonomously across ChatGPT and Slack.
• Perplexity rolled out a Plaid integration that lets users connect bank accounts, credit cards, and loans directly to its Computer agent, turning it into a personal finance hub.

Design Tools

The design tool category had its busiest month of the year, with five major releases that effectively dissolve the line between "design tool" and "AI agent that does design."

• Adobe announced a Firefly AI Assistant. Describe the outcome you want and Firefly coordinates multi-step workflows across Photoshop, Lightroom, Premiere, Adobe Express, and Firefly itself.
• Affinity released a connector for Claude, bringing AI assistance directly into the design suite.
• Anthropic launched Claude Design. It reads your codebase and brand guide to build a persistent design system, captures elements from any live site, and packages finished designs as a handoff bundle for Claude Code or exports to Canva, PDF, PPTX, and standalone HTML. Teams can comment directly on designs for precise edits, and designers can select elements and modify them with built-in UI controls. Anthropic also introduced Claude for Creative Work, a set of new connectors for creative tools like Ableton Live, Affinity, Autodesk Fusion, Blender, and Adobe's Creative Cloud suite.
• Canva launched Canva AI 2.0, which now generates and edits at the layer level, including text, elements, and colours. CPO Cameron Adams said the new Canva Design Model was trained on "structured data, millions of designs, and the actual sequence of edits used to build them."
• Cursor added a new design mode for annotating and targeting UI elements directly in the browser, plus the ability to run Cursor on any machine and control it remotely from your phone.
• Figma released Weave (formerly Weavy), a standalone node-based generation tool that lets you incorporate different prompts and AI models for exploration, comparison, and consistent image and video generation.
• Google Stitch open-sourced the DESIGN.md spec, plus a wizard to extract one from your product or website, so any coding agent can import or export your visual identity system.
• Pencil added a Code on Canvas feature that lets you ask the agent to generate custom design elements inside Pencil, create interactive components, and produce generative art, while still maintaining full manual design control.

Enterprise

• Mistral announced Workflows, a new platform for orchestrating multi-step business processes across AI tools.
• Mozilla announced the Thunderbolt enterprise platform.

Frontier Models

The closed labs kept pushing forward while open-weight labs from China continued to close the gap on context length, coding benchmarks, and reasoning ability.

• Alibaba released Qwen 3.6-Plus with a 1M-token context window and strong coding skills, alongside 3.6-Max-Preview, which took the top spot on six different coding benchmarks.
• Anthropic released Opus 4.7. The release came with mixed reviews and notably higher token costs for typical coding tasks, requiring some prompt optimization to use efficiently.
• DeepSeek open-sourced DeepSeek-V4, a reasoning model with a 1M-token context window that approaches the performance of top-tier closed models.
• Ineffable Intelligence launched as a new lab founded by former DeepMind researcher David Silver. Based in London, the company aims to build AI that learns from experience instead of training data.
• Meta finally released Muse Spark, a long-anticipated frontier model now rolling out across Meta's suite of products.
• Mistral released Medium 3.5, a 128B model with 256K context, alongside remote coding agents in Vibe that ship GitHub PRs asynchronously and a new Le Chat Work mode for multi-step tasks.
• Moonshot AI released Kimi K2.6, a powerful open-source coding and agent model.
• OpenAI released GPT-5.5, a much-improved model by most accounts.
• xAI released Grok 4.3 with an improved architecture and a December 2025 knowledge cutoff.

Images

Image generation tools focused on customization, editing, and brand-specific training over raw generation quality this month.

• Ideogram launched two big features. Editable Text Layers lets you change the font or update the copy in your AI-generated images without re-prompting, and Custom Models let you train your own model on 15-100 captioned images to define brand-specific art direction, typography, and visual identity.
• Midjourney released v8.1 with much improved aesthetics and the ability to generate HD images. It's also much better at rendering text than previous models.
• OpenAI released ChatGPT Images 2.0, a major upgrade that can search the web for real-time information, create multiple distinct images from one prompt, and double-check its own outputs. As a result, text rendering is flawless, and it quickly topped Nano Banana on the image leaderboards.
• Recraft rolled out a major UI update for its image generation platform.

Search

• Exa released Deep Max, a new agentic search tool that tops existing rivals on accuracy while running 20x faster.
• Google released Deep Research Max, a SOTA agent that uses Gemini 3.1 Pro to generate research reports from the web, uploaded files, or any MCP server, complete with charts and infographics.

Voice and Transcription

• xAI launched Grok Voice Think Fast 1.0, a SOTA voice agent that tops speech benchmarks across the board, and is already running Starlink's phone support line.

Video and Avatars

• Alibaba officially released Happy Horse, a new SOTA video generation model that has topped the Artificial Analysis leaderboard.
• HeyGen launched Avatar 5, the latest iteration of its synthetic avatar model.
• Sync released sync-3, an updated lip-sync model.

Wearables

• Button is another AI note taker, this time inspired by the simplicity and form of the iconic iPod Shuffle. It's currently available for preorder.

World Models

• Odyssey announced an open beta for Odyssey-2-Max.
• SpAItial launched Echo-2, a new SOTA world model that turns text or photos into explorable 3D worlds, claiming to beat World Labs' Marble 1.1 across benchmarks.
• World Labs released Marble 1.1 Plus, their most advanced model for creating the largest worlds yet.

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 8.1. Editing assistance provided by Claude Opus 4.7.