The last few AI releases, probably since around Opus 5.1, had been sort of ho hum. I mean, look, it’s still pretty strange, crazy times, but using those incremental models didn’t give you the “oh my goodness, I can’t believe that worked” experience that had been the norm for the Opus 4.8-era releases.
When Opus 5.5 was released, I just happened to have a project I wanted to try. My ScanSnap ix500 scanner (which I’ve used to get paper out of my life as much as possible) has been complaining about its software for a while, and, honestly, the software for it sort of sucked anyway.
A while back, I’d asked Opus (probably around 5.1ish) to try to figure out the ScanSnap’s API so that we could build a better app. It didn’t get very far.
On a lark, I tried it again.
Holy shit.
It did some research, found some existing code that got to some of the ScanSnap’s API, and then built a test protocol to find the rest. It gave me some tcpdumps to run, analyzed the results, and within maybe 10 minutes, had a working script that could talk to the scanner over wifi, initiate a scan, and capture PDFs and images off of the device. A few minutes later, it had a working macOS app (thanks also to Apple’s new Xcode features that make working with agents much nicer).
I had basically entered a few line prompt and then ran some commands it asked me to.
So, we dug a bit deeper. I noticed the scans weren’t as “crisp” as from the ScanSnap software, so it compared them, found that depending on the profile, it was making some colors closer to white (for paper), and it was making the scanned text crisper. It took 4 or 5 samples, then built a comparable color shifting model.
It added in OCR (which has been both slow, and is now broken, in the current version) using Apple’s vision models. Rather than taking an extra 10-30s, it takes 1-3s.
Over the course of a few hours, as a background task, I simply just asked it to match (or make better) features. Over the course of a day, mostly with me doing real work and popping over to check in every now and then, I had built a full replacement for my ScanSnap software, with all the features I care about (plus some new ones), with no external dependencies.
As I’ve thought about it more, and done some more work, I really think what has changed may be less the model and more how the Claude Code harness works. It seems like Claude is much better now at figuring out the right tools to use, iterating through outputs, driving itself towards a result, and doing it pretty efficiently. Things that previously required a lot of user interaction, now just prompt you when they are things Claude can’t do inside its sandbox.
The other amazing thing on this build was I never ran out of tokens. I’m on the $20/plan, not some crazy AI maximalist feeding it how often I pooped today. Still, had no issue building a full app with all the bells and whistles.
I don’t think I’ve felt this way about technology since getting on the internet back in the mid-90s and seeing a new bit of tech blooming every few months, or finding a new thing I didn’t know about.