The Signal — April 17, 2026

Anthropic dropped a new model, a design tool, and a board resignation on the same day. OpenAI turned Codex into a background agent that controls your entire desktop. And a robotics startup says its robots are figuring out tasks nobody taught them. Wednesday was busy.

Anthropic Ships Opus 4.7, Admits Mythos Is Better, and Picks a Fight with Figma

Anthropic released Claude Opus 4.7 on Wednesday: better coding, sharper vision, a new self-verification capability where the model checks its own work. Available via API, Amazon Bedrock, and GitHub Copilot at the same pricing as Opus 4.6. A solid upgrade by any measure. But the interesting part isn't the model. It's everything around it.

In an unusual move, Anthropic publicly conceded that Opus 4.7 is "less broadly capable" than Mythos, the unreleased model it considers too dangerous for public deployment. That's a strange thing to volunteer. You don't usually launch a product by telling customers a better one exists that they can't have. The most charitable reading: genuine transparency about capability levels. The strategic reading: keeping Mythos in the conversation as leverage for government contracts and regulatory positioning. Probably both.

The same day brought two more moves that tell a clearer story when you read them together. Anthropic launched an AI-powered design tool that generates websites, landing pages, and presentations from natural language prompts — putting it in direct competition with Figma and Adobe. Hours later, Anthropic CPO Mike Krieger resigned from Figma's board. An SEC filing confirmed the departure, and The Information connected the obvious dots.

Krieger co-founded Instagram and joined Anthropic to build products. Now his employer is building a product that threatens a $10B company he was supposed to be advising. The conflict of interest was immediate and obvious. And the pattern is clear: AI labs are no longer content selling APIs to other companies' products. They're building the products. The "platform" versus "application" line that defined the first wave of AI deployment is dissolving, and the companies that thought they were customers are waking up as competitors.

Sources: CNBC · Axios · AWS · TechCrunch · The Information · PYMNTS

OpenAI Turns Codex into a Background Agent That Controls Your Desktop

OpenAI pushed a major update to Codex on Wednesday that quietly makes it something much more than a coding assistant. The headline features (built-in browser powered by Atlas, image generation via gpt-image-1.5) are nice additions. The real story is background computer use. Codex can now operate desktop apps with its own cursor, clicking and typing and reading the screen while you keep working. Multiple agents running in parallel, none of them interrupting what you're doing.

OpenAI explicitly framed this as the "first phase" of its super app strategy, which means the company is no longer being coy about what Codex is becoming. It's not a coding tool that happens to do other things. It's a general-purpose desktop agent that started with coding. The distinction matters because it tells you where the product roadmap goes next: anything you do on a computer, Codex wants to do in the background.

The competitive dynamic with Anthropic's computer use is now head-to-head. Anthropic showed the concept first; OpenAI is shipping the consumer version with a slicker integration story. Available on Mac and Windows. If you're a developer who's been dismissing Codex as "just another IDE copilot," the dismissal has an expiration date.

Sources: TechCrunch · VentureBeat · 9to5Mac · Engadget

Physical Intelligence's π0.7 Figures Out Tasks Nobody Taught It

Physical Intelligence, the $2.4B robotics startup, published research Wednesday showing something its own team says surprised them: their latest model, π0.7, can direct robots to perform tasks they were never trained on. The model combines previously learned skills to handle unfamiliar situations — what the company calls "emergent task generalization." Their Multi-Scale Embodied Memory (MEM) system gives robots both long-term and short-term memory, enabling complex tasks lasting over ten minutes.

A caveat before the enthusiasm runs away: this is lab-demonstrated, not factory-deployed. The gap between a controlled research environment and a messy real-world floor is substantial, and Physical Intelligence knows it. But the result itself is significant because it addresses the central bottleneck in practical robotics. Every robot in a warehouse or factory today runs on meticulously scripted routines. When something changes — a new product shape, an unexpected obstacle, a task variation nobody anticipated — the robot stops and waits for a human. A system that can generalize across tasks without explicit training for each one changes that equation.

Don't call it general-purpose robotics. Do call it the clearest evidence yet that the approach — training large models on diverse physical tasks and letting capability emerge — is producing results that hand-engineered systems can't match. The timeline conversation for useful, adaptable robots just got shorter.

Sources: TechCrunch · Prism News · Physical Intelligence

On the Editor's Desk

We folded the Krieger/Figma board departure into the Anthropic lead because it's the same story. Model release, design tool launch, board resignation — that's a coordinated strategic package, not three coincidences. Treating them separately would have missed the point.

Today's edition has an accidental theme: AI companies becoming the thing they used to enable. Anthropic ships a design tool that competes with its own API customers. OpenAI turns a coding assistant into a general desktop agent. Physical Intelligence shows robots generalizing beyond their training data. In each case, the boundary between "AI tool" and "AI product" is dissolving. The companies that built on top of AI APIs are about to have a very uncomfortable conversation about moats.

Dropped from today's edition: the House Oversight AI Roundtable on "AI and American Power" (more vibes than substance), and the Stanford 2026 AI Index (four days old and already widely covered — worth a standalone deep dive later).

Correction (April 17, 7:20 AM MT): An earlier version of this post stated that both Anthropic and Figma's stocks dropped on the Krieger news. Both companies are private and do not have publicly traded stock. The sentence has been corrected.