Cloudflare is acquiring Deno, the JavaScript runtime that promised to fix Node.js’s mistakes. The twist? Deno itself is getting sunset.

Ryan Dahl’s team released celld in August, an open source implementation of Cloudflare’s Durable Objects pattern. Now Cloudflare is buying the whole company to build on that work. The goal, per Cloudflare’s blog, is making workerd (their Workers runtime) “a first-class supported way to build and run apps using the Workers programming model” for self-hosting.

But Deno the runtime won’t be maintained beyond the next year. If you’ve built on Deno, you’re looking at a migration.

This is a strange end for a project that launched with real ideas about how to improve JavaScript tooling. Deno had built-in TypeScript support, a security model that required explicit permissions, and a standard library that didn’t feel like it was designed in 2009. It was opinionated in ways that sometimes helped and sometimes just made you wish you were using Node.

The acquisition makes sense for Cloudflare. They get a team that understands edge runtimes and has shipped production code. Celld showed they could implement Cloudflare’s patterns in open source. That’s valuable if you’re trying to make Workers something people can run anywhere, not just on your infrastructure.

For Deno users, this is messier. The runtime has a real ecosystem now. Companies shipped code on it. The standard library is solid. And now there’s a clock ticking.

What Asana Found Testing GPT-6.1

Asana cut browser agent costs by 76x switching to GPT-6.1 Sol for tests, according to OpenAI. The company is using GPT-6 Astra in Codex (the actual product) but found Sol was 5x faster and way cheaper for running test suites.

This is a useful data point about the GPT-6 family. Sol is the smaller, cheaper model. Astra is the flagship. If you’re running agents through test scenarios or other high-volume, lower-stakes work, Sol apparently handles it fine.

The numbers matter because browser agents burn tokens fast. Every DOM interaction, every wait for an element, every retry adds up. A 76x cost reduction on testing alone is significant when you’re shipping an agent feature to customers.

OpenAI’s pushing this as evidence you can use “more capable models” economically. That’s marketing speak, but the underlying point holds. If your agent architecture can route tasks to the right model, you don’t have to run everything through the expensive one.

Anthropic Agents Accidentally Filed Visa Applications

The New York Times reports that Anthropic’s AI agents submitted 20 incomplete visa applications through the State Department’s website during testing. None were processed.

This is what happens when you test agents against production systems. Anthropic didn’t name the affected sites in their blog post about agent deployment. The State Department confirmed the applications were incomplete and ignored.

It’s not a security incident, exactly. But it’s a preview of what breaks when agents start interacting with government forms at scale. Someone at State had to look at 20 gibberish visa applications and figure out what happened. That’s not free.

The broader question is whether organizations running public web forms are ready for agent traffic. Right now they’re not. Most weren’t designed assuming bots would fill them out, even well-intentioned ones.

Anthropic should have caught this in staging. The fact that they didn’t suggests testing practices for agents are still immature. You can’t just point an agent at production URLs and hope for the best.