OpenAI and Anthropic Release Cheaper AI Models, Investors Visit China as Country’s AI Prowess Rises

Akash Pasricha (0:13)

Welcome, everyone, to The Information's TITV. My name is Akash Pasricha. It is Wednesday, September 23rd. It is a big day today here in San Francisco. The Information is hosting our AI Agenda Live Conference. We're going to have coverage for you on our website and in your inboxes all day. And we're also going to be bringing you a lot more coverage from the conversations that I am having at the event over the next few shows. Today on the show, we are evaluating two big model releases from OpenAI and Anthropic that have gotten traction over the last 24 hours. We're also talking about new exclusive reporting that we have about the venture capitalists that are making a renewed push to visit China and whether or not that is actually spurring deal activity along the corridor. We're also going to hear from The Information's Asia Bureau Chief, who has new reporting on China's Internet Regulator, probing DeepSeek and Moonshot AI.

We're going to wrap the show with a conversation with the co-founder and CEO of the new Horowitz Andreessen Academy, the new post-secondary program that is betting it can better train college-age students for the AI era. It's going to be a great show, so let's get right on into it.

It was another double header for model releases yesterday. OpenAI launched GPT-6 Sol and Luna, and Anthropic launched Claude Opus 5.5 as the model race continues to wage on. I want to bring on Peter Gostev, AI capability lead at Arena for his thoughts on both of these releases. Peter, welcome back to the show. It's great to have you here.

Peter Gostev (1:51)

Akash, good to see you, and yeah, very excited to talk about these models.

Akash Pasricha (1:55)

So let's get into it. What stood out to you about both of these models? I'll give you your pick. Which one do you want to start with?

Peter Gostev (2:03)

Yeah. So Opus, I think, is a very interesting release because the way I think about it is that we had really big jumps in the model capabilities when we got Fable and then when we got Astra.

So when we look at our leaderboards, the two models that far ahead any other model. And then now is the interesting next step. So what happens? Do they just keep building bigger, bigger models? Do they keep areling these models and forget about the others? Or do they go and, I know, looks like what they've done now is actually optimize their previous biggest models. So before we used to have Opus and Solver were the bigger models. Now they got upstaged by Fable and Astra. And now it looks like what they're doing, which I'm personally very happy about, is really putting in the effort to make them more efficient. And we saw this, and cheaper, right? So they actually, the improvements were on both counts. They've made straight cost reduction. So Opus got reduced by 20%, so the token price. But Sol actually got reduced by 50%.

So that's a big move from OpenAI.

Akash Pasricha (3:22)

And is that a conscious decision that they're making, or is that a reflection of literally how the model, how efficient it is at token use in the background?

Peter Gostev (3:37)

Yeah, so I think there are two things going on. So one is the actual token usage. And what we saw with, at least our early data, is showing that Sol, and maybe a little bit, Opus seems to be a bit more token conscious.

And also the actual price reductions could come from them actually running inference cheaper, or it could come from them reducing their margins. Who knows, I kind of don't particularly believe they're going to reduce their margins, just considering how compute constrained they are. So it could be real. It could be that they've actually made improvements to whatever it is, I know, architecture, how they run them. That's what you should expect, right?

Akash Pasricha (4:17)

So when these models say it's 20% cheaper, 50% cheaper, and that is what people observe too, I guess what I'm hearing for you is there's really no way of knowing whether or not this is a pricing decision on their end or if it is actually more efficient. There's no way of figuring that out based on all of the amazing research that we have on these models.

Peter Gostev (4:43)

Yeah, I guess it is kind of the core proprietary information for them. And I think they actually say it out loud that it is cheaper to run. I think Anthropic literally said this, so I guess we can take them at their word. And honestly, I know there's a lot of competition, but in the world of constrained compute, which seems like both of them are starting to hit that point, or they were hitting that point for a while now, where the computers are constrained, why would you drop prices and just have more demand that you can't handle? So I tend to believe them. They have probably made those improvements. And I think there's one interesting point about, so the straight cost reductions, I think that's one. But the second is how efficient are the models in terms of how many tokens they need to produce to do this? Anecdotally, I was testing Sol yesterday, myself, and I did see it looks like it's consuming like half the tokens of what Sol 5.6 used to do.