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

0:13Welcome everyone to the information's TI TV. 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.

1:00We'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. And 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 Soul and Luna, and Anthropic launched Claude Opus 5.5 as the model race continues to wage on.

1:41I want to bring on Peter Gustave, 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. Kash, good to see you, and yeah, very excited to talk about these models. 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 yeah so i and opus i think is a very interesting release because the the way i think about it is that we had really big jumps in the model capabilities when we got fable and then we when we got astra right so when we look at our leaderboards the the two models that far ahead any other model and then now what it is the interesting next step so what happens do they just keep building bigger bigger models do they keep like are rallying 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.

2:56And 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 actually, the improvements were on both counts. They've made straight cost reductions. 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. Right. 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? Yeah. So I think there are two things going on. So one is the actual token usage.

3:41And 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 don't know, architecture, how they run them. That's what you should expect, right? 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.

4:36There's no way of figuring that out based on all of the amazing research that we have on these models. Yeah, I guess it is kind of the core proprietary information for them. uh and i think they actually say it out loud that it is cheaper to run i think uh anthropically chile said this so i guess we can take it uh take them at their word and honestly i i know there's a lot of competition but in the in the world of constrained compute which seems like they're 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 they just have more demand that you can't handle.

5:11So 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 sol56 used to do um so this is a really nice trend and that it could actually deliver the same or better results for fewer tokens they've made it 50 cheaper on a per token basis and it's consuming 50 less tokens yeah look maybe the the 50 less tokens i want to see the actual aggregating numbers this is a little bit more kind of my anecdotal view but it seems to be seems to be true so it seems to be going in that direction and and the interesting thing about this is that it was kind of the the dirty secret a little bit of many models is that it was relatively easier to hit the benchmark numbers if you're just going to output a lot of tokens so sometimes you see for smaller models you see really high benchmark score and then you actually look at the numbers or how much what the cost was or how many tokens it took to produce and it's a astronomical is like five times bigger than maybe the top tier models.

6:39And it's kind of on the one hand, you can say, oh, well, we'll do okay. You know, it's a cheaper model. Maybe it's fine. But then it kind of comes with a lot of hidden costs. It takes a lot longer to actually get the answer, right? So you don't want that. Then while you technically get a high number on the score, I still don't particularly believe the models are super good at managing their contacts. So you start accumulating that kind of baggage. So a lot of benefits come from the fact that you can actually just have fewer tokens output. That's the trend we want. And that's what we saw. And that's what we saw with Fable and Astra.

7:16And so remind me, how do these models then compare to Fable and Astra in terms of who they're meant for and what they can do? Yeah. So it is interesting. when you look at the core benchmarks that they released and we saw some early benchmarks come out they kind of well number go up you know every time all of the models are better than the previous ones i think in practice fable and astra are the top tier models that will be most expensive so when i look at our arenas agent mode um approximately uh for our median task that people do on on the arena uh fable cost five dollars to complete astra cost four dollars to complete uh then opus five was i think about somewhere like two dollars and uh sol 56 was a dollar so roughly to give you an idea now when when i look at the early numbers we haven't released them yet but it looks like uh gpt6 soul comes out at about half five six soul in terms of in terms of the cost so we see that kind of reduction so if you compare for example uh gpt-6 sol to astra it's about fifth of the cost or on an average task so and that that's why it is i think it was a bit frustrating for people to just have these super expensive models that they all want to use and then they burn their subscription in 30 minutes so now it looks to me to your question about who is it for i think opus and soul probably are the daily drivers for most of us to just use day-to-day and then you might want to upgrade for for kind of that deeper level of of understanding so peter these are the first models that we've seen released by these two companies since the pacing the frontier conversation began about i want to say maybe two weeks ago a little a little more than two weeks ago.

9:27What do you think, I mean, you know, is there any evidence that these models were addressing that concern at all in your view based on what you're seeing? Yeah, I think I did see that take, you know, is this like pacing the frontier, releasing the new great model, you know, a few weeks later. I think it actually doesn't go against that at all because I don't think this is kind of the new big frontier. And we did see some uplifting capabilities, but this is not a new frontier in any kind of big, meaningful way. They just probably RL these models a little bit more, tune them a bit better, and that is about it.

10:08So I think this is totally consistent. We're just making models incrementally better. Now, if you remember, I think it was about a week ago or two weeks ago when OpenAI solved the Millennium Prize problem Navier-Stokes with their unreleased model, you might remember they had a chart there, which is counting the success rate on some discoveries of math problems that they solved that were open. And I remember clearly the chart when they had Astra barely solving anything with no matter how much you compute throw at it. and then they had their own unreleased model that was solving, you know, multiples of that.

10:53And that seems to me that's the kind of model that will genuinely push the frontier, right? It's not, oh, it's a little bit higher, two percentage points high on this benchmark, or maybe it's index score is high. No, that's not really the point. The point is when you get the new frontier, you're going to feel it, right? It's going to be very obvious. we're going to get some new math problems being solved our heads are going to get you know exploded from all of the cool stuff they can do right for now that's not it this is just the next iteration the last question for you peter so i mean now once again openai and anthropic have released their latest models uh what are we what's what are the rumblings around what we're expecting from from Meta and from Grok.

11:42What do you expect to see from them? Well, we actually also just had releases from Grok. We had just also released from Meta. And I think Meta is innovating a lot on the product side, which is exciting to see. Both of their models are not yet at the frontier, but they're good models. And I'm sure they'll just keep getting better. We heard a lot of rumors from Chinese companies also preparing their next releases. so probably something like Kimi K 3.1 and Quan 4 and a bunch of others I guess what I think Kimi K 4 now I guess we're waiting I think the rumor was for 3.1 so I guess we can't expect probably huge jumps I mean Quan 4 I guess they got up to 3.8 so maybe they look okay let's just change the number so I guess we'll see what we get but um and I think that's that's a really interesting question that what is the response from the Chinese companies because Kimmy K3 was a really really hyped and impressive release people really love this model but now it is kind of getting a little bit outdated so we'll we'll see how whether for example with K3.1 could actually close the gap or not because right now the gap is is really expanding so yeah the the response is very interesting to look for.

13:01Great. Well, Peter, I want to thank you for coming on. That is Peter Gustave, AI Capability Lead at Arena here on TITV. The information has new exclusive reporting on a growing number of prominent VCs that are paying visits to China to learn more about the country's AI landscape. My colleagues, Julia Hornstein, Jingyang, Jira Osawa, and Rocket Drew reported that story. I want to bring on Julia to share more about what they learned. Julia, welcome back to the show. It's great to have you here. Thanks for having me. So what did you learn about investor interest in China and how Silicon Valley is thinking about that right now?

13:39So U.S. venture deals in Chinese companies are down 80 % from its peak in 2021. That's in part due to Biden-era regulations on investing in China, as well as just general political antipathy towards the region, but that hasn't stopped US venture firms from making week-long fact-finding visits to the region. These investors are largely drawn by the rapid rise of competitive Chinese AI labs, as well as their portfolio's dependence on hardware supply chains in the region, rather than the prospect of investing. And so who exactly is making these trips to China? So we reported that investors from Founders Fund, Coastal Ventures, as well as Dimension Capital Management have all made recent visits to the region.

14:33We also have reported that a partner from Coastal Noa Ventures went as well. So this is kind of interesting because, I mean, Founders Fund and, I mean, look, Vinod has, he's been very outspoken about the competition between the U.S. and China, but Founders Fund as well. I mean, these are firms that, from what I recall, they've sort of spoken out against investments in China in some capacity, right? Definitely. I mean, Founders Fund partners have long expressed China hawkishness. Some criticized Benchmark's decision to back the Chinese startup behind AI Lab Manus last year. Another partner at the firm recently criticized American tech companies for shying away from working with the government because he believes that doing so has only made China more powerful.

15:24And of course, Peter Thiel, who started the firm, has long made public statements about China. Most recently at the Aspen Ideas Festival earlier this summer, he claimed the new pope was, quote, working for the Chinese communists for calling for AI regulation. So the firm definitely hasn't shied away from expressing its views on China. So what is happening on these trips then? Are they actually scouting any investments? Are they investing at all? No, I mean, our reporting is such that no investments have come from any of these trips, mostly because of the reasons that I outlined earlier. But, you know, partners are gathering intel on Chinese startups, how far along they are, what they're prioritizing, etc.

16:09I mean, they're also trying to evaluate how reliant their own portfolio companies are on Chinese manufacturing and Chinese supply chains, which can be a big hang up and due diligence for investing in even American companies. So I think that the purpose of this trip is to really understand the interplay between their American portfolios and what's going on technologically in the region. Which companies specifically are they visiting in China? So we reported that Founders Fund visited AI startup Minimax, which I think highlights an interest in firms in Silicon Valley really wanting to better understand open-weight models.

16:50We also reported that Khosla visited Jalmi and robotics maker Unitree, which I think, again, echoes this interest in learning more about manufacturing and supply chains. And overall, I mean, these visits kind of demonstrate that investors consider China to be a dominant force in, you know, humanoid robotics, hardware, AI, so dominant that they're willing to make the, you know, trek overseas. It's kind of funny to me that these companies would open their doors to these VCs at all. I mean, like, you know, it'd be interesting to be a fly on the wall for those discussions because, I mean, you could just make the argument that a VC's job is to back companies that would disrupt incumbents, right?

17:43And so, I don't know. Like, it's kind of interesting that they open their doors to venture capitalists at all. I mean, I think it's interesting, though, some of our sources have even said that, you know, as China's become such a formidable peer, perhaps, you know, limited partners across the world may be wondering why their investors are shying away from the region, because there seems to be a lot of opportunity for, you know, at least for growth there. So I think that, you know, maybe even though the regulatory and just general political environment is such that most firms are shying away from investing right now, it could be the case that, you know, these Chinese companies are willing to keep that door open.

18:31Right, right. And maybe it's sort of an acknowledgement that maybe they're always looking for founders to back and companies to back, whether or not they back them this year or three years from now when geopolitical tensions are slightly less or different. I mean, maybe that's the thing to watch for. I mean, on that point, you started with that stat about U.S. investments in China and how that's trending. Is there any reporting that you have to suggest that maybe VCs are becoming more eager to invest in China or that maybe we can expect more of that to come? I don't know if we can necessarily say that we'll expect it to completely rebound or even expect an uptick because, I mean, time will only tell what the regulatory environment will look like.

19:24But I think that there's, you know, definitely an openness to, at the very least, as we've seen through these trips, learning about the tech environment in China. You know, some Chinese investors have called on U.S. investors to even return to the region, noting that, you know, American LPs may fear that they're missing out on a great opportunity, as they said earlier. So, I mean, who knows? But I think that at the very least, it shows that, you know, American firms are taking Chinese companies very seriously. Great. Well, Julia, I want to thank you for coming on. That is Julia Hornstein, our venture capital reporter here at The Information.

20:06The Information has exclusive reporting on a new investigation that China's internet regulator is pursuing into DeepSeek and Moonshot AI. The information is Chen Liu and Jing Yang reported the story. Here is a clip of Jing sharing more about their reporting. Hi, this is Jing Yang. I'm the Asia Bureau Chief at Information. Yesterday, my colleague Chen Liu and I reported that China's internet regulator, the Cyberspace Administration of China, is investigating some of the Chinese AI labs for potential data leaks to Anthropic. The investigation was prompted by a report that Anthropic released earlier this month in which, you know, Anthropic alleged that, you know, a number of Chinese AI firms have been distilling, extracting capabilities from Anthropic's cloud models in a practice called distillation.

21:08And while this is not the first time that Anthropic laid out this kind of accusation, it was the first time that the cloud maker actually provided extensive evidence and data to support its claims in this report that is more than 150 pages long. And specifically, Anthropic said that Moonshot and DeepSeek, two up-and-coming Chinese AI stars, have been relaying or rerouting customer inquiries, customer data to the cloud models. And sensitive government, military, or state-owned companies' data have been rerouted to Anthropic as a result. So the CAC, the Chinese Internet Regulator, immediately summoned all the seven companies named in the Anthropik report and then quickly zeroed in on DeepSeq and Moonshaw, likely because of the specific examples Anthropik provided concerning these two companies in this report.

22:21For example, Anthropik said that a user that is assessed as likely affiliated with the People's Liberation Army asked Moonshot's Kimi model to analyze surveillance footage tracking a person across hundreds of police cameras in a city in China, including cameras outside the PLA facilities and the defense-linked institutions. And, you know, again, this is based on Anthropik's allegations. And also that, you know, Anthropik said that Moonshot and DeepSeek forwarded users' requests, including, you know, sensitive data such as police security camera footage, without their own customer, their own user's knowledge, and passed over the answers to their customers as if, you know, these were processed by their own models, when in fact they were processed by Claude.

23:23And then we also reported that the investigation is still ongoing. There is no, you know, it's too soon to tell whether there's going to be any kind of penalties slapped on these companies. Authorities are still looking into extent of the breaches and also the intention behind them. But still, you know, Anthropik's report and the probe that it prompted out of China came at quite a sensitive time for Beijing. Because as we know, Chinese leader Xi Jinping is due to visit Washington and have a summit with President Trump later this week. And then we also expect the two leaders to discuss something that has been described as an AI dialogue, which is basically a mechanism, a bilateral mechanism under which the two countries would flag to each other any kind of AI-related national security incidents to one another.

24:27So the story is developing. And also today, we saw that the share prices of Minimax and Zhipu, also known as Z.ai, both of which are listed in Hong Kong, actually, you know, their share prices actually fell quite a bit, which analysts attributed it to as sort of a contingent risk coming from a reporting of this probe. from the makers of andreessen horowitz the horowitz andreessen academy is a post-secondary program that is meant to give students bleeding edge ai training through courses and co-ops and access to compute credits gagan biani the co-founder of education and training companies udemy and maven is ceo of the new school i want to bring on gagan for a conversation

25:19Gagan Biyani:welcome to the show gagan it's great to have you here thanks for having me akak so uh Tell me about this new program. How's it going to work? Well, we're launching with a one-year founding class fellowship. The idea is what if you redesign post-secondary school entirely for the new era of AI? And I think we have three core tenets. The first tenet is pursuits. This is the number one difference between this and any other school. Students will be able to choose their own paths. And we will design a system that allows them to make these paths super ambitious. These paths, these pursuits can be projects.

26:00Gagan Biyani:They can be research. They can be, you know, art or science. They can be in any subject area. And students will really be able to control their own destiny, which is a challenge in and of itself. Right. And then the second thing is we'll have courses from leading experts in the field. And third, we'll have co-op programs where they can work in companies and get real world experience. So and the first cohort, I mean, I saw two dates on there, September 2027 and then 28. So walk me through the timeline here. Yeah. So as of today, you can now apply for the founding class fellowship. It's kind of a unique thing.

26:42Gagan Biyani:It's our first year's program. We're starting with just one year tuition free starting in September, 2027. Okay. And our plan is to submit an application for regulatory approval to do a two-year program that costs money. So in 2020. How many kids will you take in 2027? 50. 50. Okay. So why does it take so, I mean, why not start this sooner in, you know, January? Well, we wanted to follow the college admission cycle. And also, building a school is a huge endeavor, right? There's physical space. You built two of them. I built many more, actually, that don't get listed on my bio. But this is going to be a full multi-year school.

27:37Gagan Biyani:You want to be thoughtful about it. And also, we want students who are coming right out of high school and you graduate in high school in uh the spring and so you start in the fall right tell me i mean you know people might position this as like a um a competitor to the teal fellowship do you see it that way not really i mean the teal fellowship is 15 last i heard it's 15 people um i think what they did is incredible i mean what an amazing gift to the world to show us, to first of all, introduce us to these amazing young minds and also to show us what's possible. But this is intended to be an actual scaled school.

28:16Gagan Biyani:It's structured. There will be a cohort of people. There will be a huge network that we will have from day one. This, I think, is, and it's a separate company. It's not a nonprofit. So I think there's a lot of things that this does that are quite different from the Teal Fellowship. And is the goal here for Andreessen to back the companies that may emerge from this program? Because, I mean, I've been seeing what you've been saying online. I mean, the idea here is not necessarily to mandate that people start companies. It might be that people decide that working for a tech company is the best alternative here.

28:57But I mean, is the idea that Andreessen may look to fund some of these companies that may come out of it?

29:05Gagan Biyani:I mean, one of the core tenets of the school is like, it's open, right? We're not trying to create something that's just for Andreessen Horowitz portfolio companies or for Andreessen Horowitz investors. The idea is that we're building something that's for everybody. You know, using the Teal Fellowship as an example, some of the biggest companies to come out of the Teal Fellowship were not backed by Founders Fund, right? Right. And so our my job is not to serve A16C. I'm CEO of a new company that has its own mandate. My job is to serve the students. So I hope there are lots of investors who will back students out of this program.

29:41Gagan Biyani:And I hope that lots of the students will make the decision that they don't have the right idea yet. And they're going to go and work at a company for a couple of years or even maybe build a career out of doing that. I feel like one of the big challenges right now for young people who are really ambitious is they're told that they should or they feel like they should start a company or bust. And that's not a message we want to set. Like some of the most amazing people in the world are working inside Stripe, Coinbase, Anthropic, you know, OpenAI, etc. et cetera. What did you learn from building Udemy and Maven and I guess all these other schools that perhaps don't get as much attention as that?

30:26What did you learn or what are you trying to do differently with this program compared to building those educational institutions? Well, since - Besides the curriculum, the curriculum is different, so I get that. But you know, building a school, I wonder.

30:44Gagan Biyani:Yeah, totally. I mean, there's a few ingredients to a school. The most important thing is a network. And so one of the things we did try to do with Udemy, which we did do with Maven, was we are, and what we're doing even more with the Horwitz and Beeson Academy, is providing a social infrastructure so that students can meet both the best peers. Like we think that this is going to attract the most cracked young kids today. And so you're going to be working alongside some people who are really going to push you. And then the other thing is we're building a network beyond your cohort. So, you know, we launched with like, I think there's like 300 people on the website, roughly.

31:29Gagan Biyani:They're all incredibly impressive. We have mentors, we have faculty members, we have guest speakers, people like Satya Nadella and Jensen Huang. And then we have instructors like Max Mullen and Steven Sanofsky. And then we have tons of amazing mentors. We're building this as a network from the ground up. And I think that's really a valuable thing. And I think that value will show itself when students come in with pursuits that they have, where they're like, I really need access to this type of knowledge or this type of compute or this type of infrastructure. And we're like, well, search the network and see if there's someone there who can help.

32:10Why not build this alongside an existing college or university? Why build this as a standalone company on its own?

32:20Gagan Biyani:Look, different people should do different things, right? I'm a huge admirer of 2U and Emeritus and other companies like that that have worked with universities. For me, I am building a bespoke experience that is extremely targeted, and I think it's going to be perfect for the young builders who want to come and push themselves. And that means that we can't compromise with another organization when building. We need to be able to have the freedom to innovate. And so with Udemy and Maven, I've always taken the alternate route because I want to be at the bleeding edge of what education is doing.