Most people in the world still haven't experienced the magic of a coding agent.
This is just an unprecedented market. Like the market for intelligence has not existed before. Everybody should be thinking in a positive-sum growth mentality.
It's huge in developing countries like Indonesia is 4% of your traffic, Brazil is 5% of your traffic, places like Vietnam, you know, like places where a $200 a month Claude Code subscription is like very expensive.
You really know you have product-market fit when like enterprises are bugging you to sign the security agreement so they can use your product.
They're just like, "Please do this because I don't know who you guys are, but a bunch of us are using this. We want, you know, everybody in the world to have that aha moment with a coding agent."
Welcome back to another episode of the Lightcone. Uh, Gary's out traveling today and will be back next episode. Our guest today is Jay V, founder and CEO of OpenCode, an open-source alternative to Claude Code that works with any model you want. OpenCode has been growing at an astounding rate this year. They're now at 4.6 million weekly active users, which is actually pretty close to Codex's numbers. Today we're going to talk about what's driving their wild growth and also Jay's winding road to get here since the company went through YC back in 2021. Jay, thanks so much for being here.
Thank you. Thank you for having me.
Why don't we just start with kind of like the crazy scale you guys are at? Um, maybe tell us about any stats you can share with us.
Yeah, yeah, I mean you mentioned the weekly actives or monthly actives. I think we ended June with around 13 million or so. Um, and that's around a 20x increase since the beginning of the year. We also recently started processing around 7 trillion tokens per day. Uh, for context, uh, OpenRouter does a total of around 6 trillion. I think again sort of beginning of the year we're probably at around 300 billion or so. Um, and then in terms of sort of our revenue off of our subscription product and if you were to pay per token with our inference and you took let's say June's data and extrapolated for the year, that would be around 31 to 33 million or so. And if you took last week's data and extrapolated for the year, that's around 38 to almost 40 million. And that's the inference part of our business or the way we sort of make money. We had launched that call it end of September, early October last year. So about 8 months to getting to around 40 million or so. I guess to sort of round out the numbers, our subscribers, so these are people that pay for a monthly subscription with OpenCode. We launched that product early March, I think the end of February, and that's grown to around 160,000 monthly subscribers. Yeah, and around that that accounts for about 18 million sort of the annualized revenue.
I saw a tweet from I think he is Thibault. I think he's a lead engineer, at least one of the main engineers on Codex, saying that something like I guess 5% of all Codex subscribers choose OpenCode as like the main harness to actually use the underlying API.
Yeah, yeah. So Codex officially, you know, supported OpenCode. What that basically means is that you can use Codex subscription in OpenCode. And yeah, and a bunch of their users use OpenCode directly to take advantage of their subscription plan.
And I think they did that right after kind of some of the back and forth you had maybe with Anthropic and Claude Code. Like tell us about what happened and how that seemed like it really fueled your growth. It was an inflection point for you guys.
Again, to sort of contextualize this, we were at around 650,000 monthly active users to begin the year, and the first week of January, we started to sort of hear some rumblings around Anthropic trying to clamp down on people using OpenCode, but using Claude Code subscriptions on there. And this was a very sort of common way to kind of use the Claude Code subscription. And the way I think they sort of blocked it, or they were trying to block it, was if the system prompt mentioned literally the word OpenCode, they would sort of reject the request. You know, from our perspective, that sort of makes sense, you know, they're sort of subsidizing usage. That's kind of what they want to do. But of course, a lot of users weren't happy, and uh, when they'd kind of done that, what it inadvertently did was it put OpenCode and Claude Code on the same sort of pedestal. It like equated the two products in some ways, and even for the people that weren't using OpenCode at the time, you know, they sort of took notice of the fact that um, that you know, Claude Code is taking that kind of an action against OpenCode.
You got new users because people heard about you for the first time, and they're like, "Oh, Claude Code's like banning this thing, or like clamping down on this thing."
Yeah, or that it is like worth looking into, that you know, it's not just one of the other dozen or so coding agents out there, yeah.
It's funny how often this happens in startups. So this same thing happened with Instacart when Amazon bought Whole Foods. It was like, "Oh, this is like the death of Instacart." And the death of Instacart became like this meme. But the meme actually just like drove all the grocers to check out what Instacart was, and then they just went through this like explosive growth of signing up every grocer in America. So it seems like actually a lot of your growth is global, sort of across the world. Tell us a bit about that.
Yeah, so the premise of the product is that you know, most people in the world still haven't experienced the magic of a coding agent and you know, it's been almost a year and I'm sure you know, like it's sort of hard to remember maybe for you guys as well, but the first time you had this experience it's a magical experience and when we had felt that for ourselves, we sort of recognized that you know, how important that moment is in sort of tech history, I guess it comes along once every sort of generation or so and the way we sort of looked at that was let's take that to as many people in the world as possible. You know, have them sort of experience something similar because the frontier models and the frontier labs charge so much per token that it's going to be hard for a lot of people across the globe to have that experience and um, we want to make sure they had that with us in some ways.
I guess there is that a point where is it true that early on there was a like seemed like there was a big gap between the open-source models and the frontier models? Um, is that true kind of when you were first launching the product?
Yeah, for sure. I think when we had first launched it was mostly hey, you're using your Claude Code subscription with Claude Code come try that with OpenCode. And then and so this was back in June of last year and by about August September or so we started to see the first crop of open-source models and you had the sense that like oh, they're maybe 6 months behind, you know, and of course there's always been sort of a gap between them, but that was sort of the first point when people were like there's the GLMs of the world, the Kimis of the world, the MiniMaxes of the world and it seems like there is now an alternative and um, and as that started to happen, that would trigger a wave of users coming in trying out OpenCode because that was probably the only way you could try out some of these multiple models and as that gap started to shrink or the fact that the open-weight models became good enough for real work is when it sort of became viable to use OpenCode with them.
Yeah, I think I remember you maybe a few months back you were saying that um, let me get this right, that even though the open-source models are obviously cheaper to use through OpenCode, you still saw more usage from the leading frontier model except was it when Gemini 2.5 came out? That was the first time that it was equivalent or it flipped?
Yeah, that's exactly it. It was like probably 2.4. I forget the sort of exact model but yeah, this was February of this year where for a 4-week span and this had never happened in our data before and I think we're sort of fortunate to be able to see this global usage and so a lot of times we see the comparison of you know, these open-source models versus some of the frontier ones and we had noticed for the first time that a bunch of users were just using Gemini way more than they were using the Anthropic models. It was sort of Sonnet plus Opus at the time together. And um, that's the point where we were like, oh, I think we should launch a subscription product because now maybe these models actually do make sense for real work.
So speaking of that like you guys have this incredibly unique insight. Like probably the only you're the most you have the best data on just how these models are being used by engineers across the whole world and you released a bunch of this data. So maybe we could just like look through some of it and pluck out some more interesting insights.
Yeah. So this is you know, you can go to opencode.ai/data. I think we started to publish this about a month ago or so. This is basically taking all of the usage on OpenCode Go which is our subscription plan where you pay $10 to be able to use any of these open-source models and this breakdown specifically is by token volume per day across the different models. So what we sort of see here is that DeepSeek Flash is the one that is used a lot. And there's some sort of details here. I can kind of go into why that is sort of the case. But, if you just sort of look at the top three, we're seeing sort of the two DeepSeeks plus GLM, and you can kind of see all the hype that GLM has been getting sort of lately being sort of reflected here in these charts.
Which the data says otherwise as opposed to all the Twitter chatter about the GLM taking over DeepSeek. This is telling a different story.
Yeah. Yeah, it is. I can show you a different breakdown here. So, this is by unique users, and this was the thing that Arjun was sort of alluding to. We get to see actual usage data for each of the users as opposed to with maybe like an OpenRouter where you're seeing it aggregated across a bunch of services or other products even that are sort of internally using it. In this case, it's you know, each one of these is sort of an actual user. But, yeah, what's sort of fascinating is if we look at the market share graph. So, this is breaking down for each of these labs the amount of token volume they're doing per day, you know, comparing that sort of across them. You can kind of see the DeepSeek one dip around the time GLM sort of comes out, but it seems to sort of bounce back up after and we have sort of some theories around why that's sort of the case, but that's sort of an interesting fact. And then, if I go back up to the users one, this is also kind of fascinating in that it might be a little bit hard to see here, but if you look at the sort of top three unique users per model, you've got Flash at sort of 38k, and then there is DeepSeek Pro at 31k and GLM 5.2 at close to you know, 30k as well. And that's interesting because GLM is sort of on par with one of the DeepSeek models, but the fact that there are two of them makes it a little bit different. I'll sort of caveat this by saying that DeepSeek Flash being as cheap as it is allows a lot of users to extend how much they can use a coding agent because as they get closer to let's say their daily or weekly limits, they can switch over to one of these very cheap models, in this case DeepSeek Flash, so they get the rest of their work done. Again, this is very different from the way we sort of think about coding agents and LLMs here in the valley, in SF, in the west in general, so.
What are you seeing as you know, I mean we definitely are the other extreme end of where it's sort of like token maxing, throwing money at it, throwing money at it. But for the yeah, the users you see, which to be fair, does seem like there's a general vibe shift, especially in the enterprise world towards like more token budgeting. What do you see? Is that as simple as people once they approach their usage limits, they switch over to one of these models or is there more going on?
Yeah, there's a few different things. I think people do try and optimize for things. One of the reasons why it was being hosted in a way that the tokens per second was a lot higher than what you would get out of even Novus. And so it was just a drastically faster model. It felt like you were working almost in real time. And some of these models tend to exhibit those qualities, which makes it you know, characteristically a little bit different from using some of the frontier ones, for example. So there's a little bit of that, but yeah, cost is obviously a big driver. And then the other one that pops up every once in a while is, you know, people sort of get a feel that GLM 5.2, for example, is better at front-end design as compared to some of the other models. And so that ends up sort of driving some usage as well.
You also have a really interesting breakdown by geo to show like where the users are. Yeah, so who are your users? Where are they coming from?
Yeah, so again, some sort of context here is that, you know, we had launched the OpenCode Go plan to be able to serve the sort of global audience and you can kind of see that with China being kind of number one at 17%.
That's crazy. I feel like you're probably the only YC company in history that has meaningful usage in China.
It's also fascinating because a lot of these models are Chinese. So, in their situation, they're trying to use the ones that are being built in their country and OpenCode kind of gives them the choice to be able to do that. So, that's sort of interesting as well. I think the US one is actually interesting to us because when we had built this plan originally, we weren't thinking about the US. We weren't thinking about the states because, you know, as we were sort of saying, people here just throw money at it. But then now it turns out it's growing really really well in the states as well and um, and maybe yeah, that speaks to the sort of vibe shift of like maybe you should be a little bit more conscious with the tokens. But then there's the other side of this where people want to use some of these. So, when GLM 5.2 was getting popular, a lot of people are trying out our subscription plan because it's one of the options to sort of do that.
So, because it's so much cheaper, it's huge in developing countries like Indonesia is 4% of your traffic, Brazil is 5% of your traffic, places like Vietnam, you know, like places where a $200 a month Claude Code subscription is like very expensive. So, and that makes sense that, you know, sure. But what you were telling me earlier was that in addition to that, there's a lot of like large US companies that have basically unlimited budgets for tokens that are also using OpenCode. Can you talk about that and why like who's using you and how come those people are using you?
Yeah, this is what we had seen early on before some of these open-source models even sort of took off is that a lot of companies would start using OpenCode because they didn't want to be locked into using a specific model or a specific harness. So, you know, some users just wanted more choice and this was a good neutral option for them that allowed them the flexibility in the future to switch to whatever they sort of wanted to.
And I think you have a crazy stat. It was something like a dozen of the top very forward Fortune 500 companies are using you and have a significant footprint.
Yeah, it's funny. We get DMs, we get emails sort of internally about hey, you know, XYZ company here. We got a few thousand people using OpenCode now. Please don't share this publicly. But I think most of it what's sort of going on is like yeah, there's a bunch of choice there. And we can probably talk about this later, but we're very intentional with our product design. We want something that everybody uses every single day and we hold that bar fairly high and in those instances, I think that's what's sort of resonating with people as opposed to the cheaper tokens thing.
I'm very curious on a slightly different topic. There's been a shift for all these companies like you in terms of the AI token economics. Because in the old world of companies, other B2C or B2B companies, there was a lot of CAC that used to be based on ads. And now there's the question is different. It's based on token to acquire users. But also even more weird, this is thing that you're describing where there's some whales that pay for a lot of it at some point. And there might be a lot of churn, but it doesn't matter because as long as you have the power users and experts really using it, they convert these large orgs. And I think that big labs can subsidize that, right? Which is what effectively Claude Code and Codex have done. They can subsidize it, but you have a magic formula to skip all this.
The broader context here is that for you to use AI, and especially these coding agents, because of the amount of tokens they use, to use them well, you have to really sort of understand them, and this is the expert thing you're sort of talking about. And to get there is fairly expensive from a per token perspective.
Token maxing is expensive.
Right. It's very expensive, and that is a chasm that is very hard for a lot of people, a lot of companies, to cross. And so, what, you know, the Anthropic's and the OpenAI's of the world do is they subsidize it so that people are able to cross that. And then, yes, like the whale's idea here, a certain percentage of them cross it, get to that point, are spending the crazy amounts that you sort of see, and then it makes sense. The entire sort of funnel then sort of makes sense. But, when we had sort of approached this, we had thought about from a product perspective where it was, you know, we sort of talked about we want, you know, everybody in the world to have that aha moment with a coding agent. So, that's sort of our free tier. And then, because the open-source models were now cheap enough and good enough for real work, you want a subscription plan that allows them to do real work with it. And that is sort of the thing that allows somebody to buy into, okay, now I can justify spending so much more to potentially redo some of the processes within my company to take advantage of these coding agents. And that's when, hopefully, you know, some of these sort of turn into whales.
And is that Sorry, and that's what you're seeing? Like, you're seeing people come in through the free tier to try it out, and then become advocates to be like, "Oh, yeah, like, we should adopt this at our like Fortune 500 company."
Yeah, I mean, it's to the point that we'll you know, in sort of the older sort of SaaS enterprise world you would have this procurement process that a lot of them sort of go through where you know, somebody sort of reaches out and you have this whole dance that they do. In our case, the inbounds that we get are typically just like, "Hey, there are a bunch of people at this company at our company using you guys. Can you fill out the security questionnaire?" And we're just like, "Oh, I First off, I didn't know. Uh, but secondarily, give us a second."
You really know you have product-market fit when like enterprises are bugging you to sign the security agreement so they can use your product.
It's like, "Please do this because I don't know who you guys are, but a bunch of us are using this." And so it is a little bit different now, yeah.
Once you're through the procurement and admin side of things, what are the enterprises asking you for? Are they trying to pull the product in a different direction? Cuz that often tends to be something that open-source companies have to think through and be careful about.
What's interesting here, I think, is these coding agents are at the core of how an LLM does work. And so a lot of times when we get these enterprises sort of reaching out to us, it's because they're trying to figure out where else they can use it. There's the obvious one, yes, we've got a bunch of developers at our company that are using OpenCode. Just figure out how we can officially use it. And then on the flip side, it is oh, there are some non-technical people that would want to be able to use this as well. And then, you know, secondarily, there's this our product is probably going to be using a coding agent as a part of its core loop. Can we sort of use it there? So, yes, so we get some sort of pull there. The other bit of pull that we see is mostly around just managing, you know, tokens a little bit, but being like, "Hey, there are certain organizations within our company that don't need the frontier models. Can we sort of limit access there or have some more creative ways of managing token spend?" And we see pull there and so you know, there's some sort of questions around that. The stranger one we'd gotten recently was somebody wanted to be able to, you know, mostly just have really good visibility of exactly what everybody is doing at the company and then that sort of gets into questions of like, you know, is that something we want to build?
Didn't Ramp use you in sort of an interesting way?
Yeah, I think Ramp was very sort of forward-thinking. They had published a blog post. I think this was in December of last year, but they had reached out prior to that where a team within Ramp had kind of built this Slack bot that was essentially running OpenCode behind the scenes and it was just sort of incredible to see. It was mind-blowing partly because we hadn't even done that yet internally and they were showing off a use case that was definitely pointing towards the future and.
What does that mean exactly? I think it can be hard for people to get their head around cuz they think of a use like an open-source Claude Code. Like what does that mean that their Slack bot was running on OpenCode?
Yeah, so you can think of OpenCode as a two-part product. There is the UI and the application part that you sort of see and that you interact with, but then there is the agent loop, the thing that's actually doing work while calling the LLM and that is sort of behind the scenes. That's a little bit of like we call it the server. That server can be embedded separately from the UI and so in this case, they were embedding
that and running their Slack bot off of that.
YC's next batch is now taking applications. Got a startup in you? Apply at ycombinator.com/apply. It's never too early and filling out the app will level up your idea. Okay, back to the video.
Can you tell us a bit about sort of these numbers and how the unit economics work? Because there was a shocking stat that was mentioned in Dylan Patel's podcast that Anthropic is profitable and by a huge margin. In Q2, they are on track to be doing 50 billion annualized revenue and around 70% margin. But they got there. Before, the previous year was not profitable and way less. And they crossed this chasm, which it sounds like where you're heading. But you don't have to subsidize it, which is special.
Yeah, I mean, we do subsidize a little bit, right? The subscription plan that we kind of have. But yeah, we have a free tier, too. We have a free tier, yeah. And I think that's the CAC part that you were sort of talking about early on where
The new access, rather than ads, the CAC is rather than paying for ads, is paying for tokens.
Yeah, it's because that's how people sort of experience that kind of magic moment. And that's the way we sort of get them into using the product, understanding what a coding agent is. All this stuff is definitely like, you know, a little too much for a lot of people. And I think the part here in terms of the unit economics that kind of works out is if you're actually paying per token there, in the case of Anthropic and in the case of how we sort of operate as well, we're able to get, at least in our scenario, we're able to get volume discounts because of the amount of tokens that we sort of serve. And so, when you pay per token, that effectively turns into our margin. Whereas, when you're subsidizing, yeah, obviously you're sort of eating the cost there. Or in the case of the free tier, yeah, you're eating the cost as well. But, you know, you were sort of talking about this before. As you start to get more and more of these whales, those whales are paying, you know, per token. And it's directly feeding into your margins.
Yeah, the discount you're able to get by sort of aggregating the tokens is interesting because I mean it's something we've talked a lot about over the last year in particular. It's where's I mean everybody in the valley is at this point. It's where's the value going to accrue? Will it be the frontier models that are going to make all of the money and everything at the app layer is left for dust or will it go the other way? How have you thought about that? Cuz yeah, you're in an interesting spot because you actually really do own the relationship with your end user and you're sort of effectively making it easy for them to pick and choose the models that they want to run with. So just what are your thoughts on where this all plays out and how it will hopefully play out in a good way for OpenCode.
Yeah, it feels like a little bit like a marketplace where a user is able to sort of make their choice of the model that they kind of want to use and different models, different attributes, you know, different cost characteristics. And our claim here is look, we want to showcase that diversity as well as sort of possible. And that also creates an environment where the labs are aware of each other and that competition ends up being good for the consumer in this case. Whereas the flip side is if you're locked into a specific vendor, then you don't benefit from, you know, the sort of competition that would otherwise sort of come with it. And you're likely, you know, helping their maybe their margins in some ways.
Yeah, I kind of feel like your growth is a pretty decent proxy for the fact that the choices only improved over time, right?
That's exactly it. Yeah. Yeah, I think every we could probably track back to sort of our bump in monthly active users down to some sort of bump in the open source model market and the way we sort of think about this is that it's not that we're picking a winner in terms of a model lab. We're just betting the field. We just think the rest of the field is going to do well.
The next logical conclusion of this is that models are becoming commoditized utilities.
Yeah, I think what's sort of interesting here is that the market is so large that it is hard to imagine people not or model labs not picking off niches and chunks of their own in that they're specializing for specific areas or specific characteristics. Like in our case when you sort of see some of the data that we were looking at before, you've got obviously the frontier labs, you know, we sort of know them well, but when we see the success of DeepSeek, it's very clear that they have picked the cost part of the quality cost performance sort of axes and then they're basically saying, "Look, we want to be very, very good at that part." And, you know, it's hard not to imagine that happening across the board across all these characteristics.
Yeah, it feels like if you truly believe the sales pitch of the labs themselves make, which is this is just an unprecedented market, like the market for intelligence has not existed before, everybody should be thinking in a positive-sum grow-the-pie mentality, in which case like they should really want you to grow and succeed because you're end up just being like a huge customer for all of them.
Yeah, I think that's sort of true now with these open-source models, we're the largest customer for most of them.
You're the largest customer for most of the open-source models.
