Did Anthropic Break Opus 5?

**Brian** (0:00)
Hey, good mornin everybody. It is August 5th, and it is a Wednesday episode, which means it ends in 3 or 8 So I think we're at 7
3 There we are. We could do math on air. You are watching and listening to The Daily AI Show.
I am Breanne Gareth. Thank you for the studio. Hey, Gareth. We can hear a little echo with you in the studio. So can you turn your speaker off, somebody? With me in the studio today is Andy, Beth, and Gareth. Yep, it's you, Gareth, because it just went away. I don't know how that happened. We will troubleshoot live on air with you because we show up, as we are, in a given day in this year of AI.

**Andy** (0:55)
Without coordination, we're all wearing green.

**Brian** (1:01)
It's that, I don't know, after almost three years, now we're completing our third year, we'll start our fourth. After three years, maybe there's just one brain about color. We're coming into alignment.
Hey, how are you today, Andy?

**Andy** (1:17)
I'm well, thank you.
I'm a little flummoxed by the seemingly consistent commentary about Opus 5 being disappointing in various ways, because I'm not having that experience myself. So I'm questioning myself, why am I not having any trouble with Opus 5, and everybody else is complaining about Opus 5? So I'm a little confused. I'm swirling and wondering whether I should switch models or no, because it just seems to be doing an excellent job for me.

**Brian** (1:56)
I think that's fascinating, and that's something that I'm interested in talking about too today, because I think people who developed a kind of shorthand with their experience with these are having trouble, and I think the people who are like, no, I communicate clearly, these are the goals, these are the scenarios, this is the stuff, this is my PRD, this is where we are in the development process, are happy because it's able to do more within that guidance.
But I had long conversations with Opus. We're going to try Gareth again. Go ahead and unmute.

**Andy** (2:44)
No echo. You sounds good.

**Brian** (2:47)
So I can hear an echo.

**Andy** (2:49)
Oh, now I hear it. Now I hear it. Oh, weird.

**Brian** (2:52)
Is this in my head? Because now we're, no. All right, thanks, everybody.
For example, what happened with me with Opus last night? And it relates to what we've talked about. And I'm not hearing an echo anymore, Gareth. So good, whatever you just did.
It relates to conversations we've had. I wanted to know best practices. How have people solved this before? This is a conversation I have with AI, specifically Claude, a lot. And it used none of the context. Like part of it is that we've had this conversation so many times, there's a ton of history. What I mean by that, what the instructions are, all of those things, which is why I shorthand it. What are the best practices? Go do that. Opus 5 knew that I was asking about best practices, did not cull the context of the knowledge and spawned literally general open-ended agents. This process for me typically is done in five minutes, seven minutes maybe. After 20 minutes, because I could see how long they've been running and nothing's come back. I'm like, hey, this has taken an unusually long amount of time. What's happening?
Opus said, yeah, let me check on that. I said, let's stop the agents, right? Let's like, give me what you've got now. Opus said, oh, the agents that I spawned, spawned their own agents and I can't stop them. I don't have control of the sub-agents that the sub-agents spawned. And I was like, yep. And that is the kind of thing that is like, what is happening because all we did was update a model. When we went from 4.7 to 4.8, this kind of like, boop, I have no idea what you meant, it was not a part of it. When I said go and look, has something been deleted? Have the, is the learning still there? Like, what is happening? Oh, it came back because that's a research question. So if you're saying to an AI, why did you do this? You are going to get a most probable response to that if it doesn't actually run research to observe something. If it observes something, you're still getting the most probable response, but you get a little bit more in terms of, it's not just like making it up in response to that. It's actually pulling more context in.
It said, no, it all exists. I have the memory file, all of that.