Update on Claude Opus 5.5
Posted by T.Collins Logan onJust a quick note that, as of Claude Opus 5.5, I’m now relying on AI to be a collaborator across several different areas I previously could not. I do have preferences and “skill” instructions loaded that help remedy some of the previous issues I experienced, but 5.5 has definitely turned a corner. The 25% error rate is greatly reduced, as are hallucinations, and its inability to self-critique. I now use Claude’s Mac client (for local chat and user emulation), as well as Claude Code. And my previous frustrations (endless corrections, redirections, and having to push back...which resulted in a lot of wasted time and effort) are now attenuated. This could, at least in part, be due to my learning how to query more effectively. For example, one of my standing rules for Claude is that I encourage it to ask me questions if it is unsure of scope or intent, and that has been a reliable way to avoid Claude’s previously rampant rabbit-hole tendencies. But I don’t think that’s the only reason...as I watch how Claude approaches each task (i.e. its “internal dialogue”), I’m seeing real clarity around effective and efficient processes. It’s getting smarter, and more trustworthy, even on extremely complex research and coding. And I’m actually relaxing what had been a constant vigilance around Claude’s fuckups.
Does Claude still make mistakes or misunderstand instructions? Yes, sometimes pretty glaringly, but it is now a tool I feel a lot more comfortable relying upon. I just thought I’d update folks on this status, since many of my previous posts have been harshly critical of AI – though I still think they were accurate for those previous iterations.
That said, the potential dangers of AI are becoming increasingly obvious by the day; essentially, models are already powerful enough that an individual with a laptop can accomplish the same damage (to infrastructure, banking systems, industrial safety, etc.) in a few days that a state-funded hacker group would take months to engineer. And of course there are still the growing threats to everyone’s economic security as companies begin relying on AI for routine tasks, the environmental and electricity impacts of data centers, and the oligarchic influence of Big Tech. I’m also still not seeing how Anthropic will become a profitable company in any predictable amount of time. With all of this in mind, and in the very myopic considerations of my own day-to-day work, at least Anthropic’s product is closing the gap between irritating toy and effective collaborator.
My 2 cents.

