Anyone who uses Sonnet for simple tasks in a workflow just got a massive intelligence boost and a price drop for those testing, tools calling and debugging jobs. Big brain work is still an Opus 4.8/ChatGPT 5.5 task (and hopefully Fable 5 and ChatGPT 5.6 soon) but if you are into the multi agent workflow and need a smart, cheap minion to leave more tokens/$ for your larger models then this is a great day for the coding world I think. There will be a lot of hissing and booing online because its not Fable 5 but I think they are missing where the value has been added. A rising tide lifts all boats and this is definitely raising the baseline. https://www.anthropic.com/news/claude-sonnet-5

Here is a technique I use (in a recent livestream) to help your ai agent give better UI layouts for your apps. Prompt: Here’s my phone, an iPhone Pro Max. My thumb only reaches the bottom comfortably, left or right hand. Put every control the user taps in that bottom thumb zone. Use the top 65 percent as a passive zone with no controls, just the info that builds trust, like the detected input and the transcript. Make every interaction take about a second and a half, and make it so a baby could figure it out. And take out anything that isn’t pulling its weight.

https://www.rayfernando.ai/glm-52-easy-setup I wrote up a guide on my blog post so that folks can get up and running with GLM 5.2. This is probably the best way to do it inside of Droid. That way you get the Million Token Context Window and a really good harness to drive this. Another advantage of using Droid is that you can bring your own key, and in this case I explained how to do that. What's nice about my blog posts is you can also point your AI agents to set them up. I include setups for doing this inside of Claude Code, Codex, and Factory AI, but you can point it at other harnesses as well.

Istead of one agent doing everything in a line, the main agent breaks your goal into independent slices and fans them out to a whole team of workers at the same time; some read and analyze your files, some go research the web, and each one comes back with a clean little report. Then the main agent stitches all of it into one result (a roadmap, an audit, a research brief, whatever you asked for). Here is the fun part. I reverse-engineered how a top-tier model decomposed a giant prompt into parallel agents, then packaged that workflow so anyone can run it on demand. And I put it through a real run, bugs included; it read about 2,500 of my own messages, researched the entire tech stack, and handed me back a roadmap in a few minutes with 11 agents going at once. What you get: - A drop-in Cursor skill (unzip into ~/.cursor/skills/, reload, then just describe a big research, analysis, or audit task) - A built-in structured format so every agent reports back the same clean way - A handoff prompt if you want to port the same idea over to Codex This is the kind of thing that used to feel like it needed a team. It does not. The power is already yours; you just have to ask. Drop a comment with what you run it on, and let me cook with you.