← all posts
Daniel Bilsborough
Daniel Bilsborough

Opus 5.5 computer use and Claude in Chrome: AI coding agent thoughts

For the last two weeks I’ve pretty much just been using Opus 5.5 over Fable 5.1 for the majority of my work. A few things I found along the way.

Opus 5.5 is brilliant at driving a browser

If you install CUA and teach Opus how to use it, it can drive your whole computer and a browser. You can also give it access to the browser through Claude in Chrome, and then it can use Brave and Chrome at the same time, with multiple tabs on each, doing multiple things at once.

You don’t need any kind of virtual machine for this. You do need some kind of screen lock, because sometimes the agents are copying and pasting, and things get confused. I run a simple lock so only one agent uses the screen at a time.

Going back to Fable 5.1

Using Opus 5.5 only for two weeks and then going back to Fable 5.1, I noticed how much “smarter” Fable 5.1 is. After re-normalising on Opus 5.5 that consistently, Fable 5.1 is a much better thinking and strategy partner for data analysis and strategy for business growth, ads, SEO and marketing. That lines up with how I split the two in Fable 5.1 versus Opus 5.5: Fable for thinking and strategy, Opus for the coding and grunt work.

GPT-6 Astra and trust

I very briefly used GPT-6 Astra in early October, but I found it less trustworthy than all the Claude models. What I mean by that is I’d tell it to do something and sometimes it would just stop working, and never tell me that it stopped or why. I had to come back and prompt it again. Claude just gets things done, and the same goes for Grok. Grok 4.7 is very good at being a trustworthy, proactive agent compared to ChatGPT in many different ways.

Bot apps: Grok Bot, ChatGPT, OpenBot and Claude Duet

I was using Grok Bot with Grok 4.7 for a little while. The Grok Bot app is a very good app, the best bot app I’ve used, and I found it much better than ChatGPT. I’ve yet to try OpenBot, but I do plan on doing that.

I haven’t tried Claude Duet yet either, but I’ve joined the waitlist, as I do want to try it. The reason I’m not too concerned about any of them is that I’ve been using a setup that does exactly what the bot apps do, built with tmux and a CLAUDE.md file.

AI voice and video: ElevenLabs V4, Seedance 2.5 and Wan 3.0

Some other things I’ve been playing with:

  • ElevenLabs V4 is unbelievably realistic voice generation. I’ve been using it for AI video narration for some of my clients and some of my own tools.
  • Seedance 2.5 Standard is the best for video I can find so far. It’s also the most expensive, unfortunately. Seedance 2.5 Turbo is a good alternative at about a fifth of the cost most of the time.
  • Wan 3.0 I’ve experimented with a little, but it’s not better than Seedance 2.5 Standard for prompt accuracy and prompt adherence.

More detail on those in ElevenLabs V4 and Seedance 2.5 for training video and Seedance 2.5, Wan 3.0 and ChatGPT Image 2.0.

Staying at the coding agent level

I prefer to operate at the coding agent level as opposed to the desktop app level, even though the Grok Bot app is very good. I can do the same thing easily through the command line using tmux, ShellDrop and my own custom apps.

At this stage I’d rather build a fully customised business agent operating system than integrate something like Grok Bot. The phone end of that setup is in how I run Claude Code via tmux from an iPhone, and the wider picture of Claude Code is on its own page.

Daniel Bilsborough

Daniel Bilsborough is an AI advisor for founders and business owners in Australia. Full advisory engagements, implementation roadmaps, and ongoing advisory.

Full advisory is $15,000. A deep session on your business, a written roadmap, and 30 days of direct access while you put the plan into practice. It includes planning custom apps for your business.

Contact Daniel →

Daniel replies by email.