GEO Kills the Listicle

**SPEAKER_1** (0:06)
This is Marketing Over Coffee with Christopher Penn and John Wall.

**John Wall** (0:14)
Good morning, welcome to Marketing Over Coffee. I'm John Wall.

**Christopher Penn** (0:18)
I'm Christopher Penn.

**John Wall** (0:19)
And we are the number three business podcast in Malawi, Africa. So I just had to throw that out there because that's just so cool that people in another part of the world are listening to us. There's no other point to that headline but that. So in more useful news, Claude has been updated.
We have become all Claude all the time. So what's the latest coming out of them?

**Christopher Penn** (0:41)
So the new model was dropped yesterday, which would have been June 30th, as of the day of recording. It's funny. It actually dropped while I was in the middle of a client call talking about Claude and it's like, oh, I'll do like, oh, all right. Well, so that's changed everything. So there's four models in the Claude family now. There's the fifth generation Fable, which we talked about previously, which apparently just got re-approved for use by the US Department of Commerce. So that will be back shortly. There is Opus, which is at version 4.8, which is their fourth generation, very smart model, very expensive. There is Haiku version 4.5, which is their super lightweight based documentation reading model. And there is Claude Sonnet, which is sort of their mid-range model. Sonnet just got upgraded to Sonnet 5 It's the fifth generation model, the second in that family besides Fable. And I've been testing it out for the last 24 hours.
And noticed a couple of things. One, it is more token intensive than its predecessor. About 1.3, so about 30% more token intensive than Sonnet 4.6, which is the previous version. And it is agentically tuned. So basically, if you look at this, this is clearly a derivative of the other fifth gen models just made more practical workhorse. It's somewhere between Sonnet and Opus in its quality. Its speed is slower, and boy, does it make a lot of tool calls. So it uses a lot of external functions.
Anthropica said this is because it is an agent-first model. So which is an interesting thing because it really shows that the industry overall is pivoting away from chat-based models, where you're having a conversation to agent-based models, where you hand over a project plan and say, okay, you do it.

**John Wall** (2:26)
Yeah. Well, and that matches with all the Salesforce Connections stuff.
We've been running too, as far as them just going all in on everything, agentic, all the time, like you said, stepping further away from the chat. Yeah. How about as far as just the increased token usage? I mean, is that actually going to change the way you're looking at things, or it's more just a matter of like, okay, we just have to work around this?

**Christopher Penn** (2:45)
If you are used to using these tools in an agentic fashion, it's not going to change a lot because you're still following the basic recipe of use the smartest model to build the plan and have the lighterweight model execute the plan.
What's interesting is that because Sonnet is more expensive, but Haiku, which is the lowest model, is still pretty dumb. There's now this weird little gap in between where if you wanted something that was a good workhorse, but not too expensive, it's no longer squarely sits in that place anymore. Now, it's on par with just below Opus 4.8, which opens the door to using third-party models. Outside of those, if you are a Claude shop, you might want to look at something like either a GLM 5.2 or a Minimax M3. I've been using Minimax a ton, and you could even use a local model like Gemma 4 or Quen 3.6 to do the workhorse stuff.
Copy this HTML, write this code based on this plan, search the web for this. So I use Gemma 4 for deep research now, because in an agent framework, I can say, here's your 12-page research plan, just follow the plan. And it's like, okay, I'll go Google this, and it comes back an hour and a half later saying, I did all the Googling, and like, good job, you Googled. And now I can hand it back to the smart model for processing. So one of the things that people need to be really thinking about is how you structure your workflow with these tools and how you decide who gets what kind of work from the different AI agents. In a lot of ways, it's no different than delegating to different team members on your team to say like, yeah, you're gonna do this, and you'll do that, and you'll do that.