Y Combinator Analysis: The Rapid Growth of OpenCode—20x Growth Since Being Blocked by Anthropic, Global Expansion of Open-Source AI Agents, and a Focus on Developer-Centric Terminal UI
▶ Original Video: https://www.youtube.com/watch?v=_O6x4ktK6JA
July 24, 2026
Program Introduction
This episode features Jay V, CEO of the open-source coding AI agent 'OpenCode,' appearing on Y Combinator's podcast, 'Lightcone.' Please note that co-host Gary is absent due to travel and will return in the next episode. As an open-source alternative to Claude Code that can integrate with any model, OpenCode has shown phenomenal growth this year, with 4.6 million weekly active users, approaching the numbers of Codex. Today's theme is the driving force behind this rapid growth and Jay's winding journey since the company joined YC in 2021.
Overview
Jay V, CEO of the open-source coding AI agent 'OpenCode,' appeared on Y Combinator's podcast, 'Lightcone.' OpenCode has reached 4.6 million weekly active users, achieving explosive growth of approximately 20 times since the beginning of the year. Monthly active users are at 13 million, daily token processing exceeds 7 trillion, and revenue from token usage alone is nearing an annualized rate of nearly $40 million, with subscriptions bringing the total to an annualized $56–58 million (approx. 9.17–9.5 billion yen at 163.69 yen/dollar). The background to this rapid growth includes the ironic catalyst of Anthropic blocking OpenCode usage on Claude Code, which actually increased the company's visibility, and a structure where developers worldwide, especially in emerging countries, are flocking to cheaper open-source models. The episode also focuses on Jay's persistent entrepreneurial story over 16 years to reach this product.
Exploding Numbers: 4.6 Million Weekly Active Users and $58 Million Annualized Revenue
The statistics revealed by Jay show that the company is on an exponential trajectory. As of January 2026, monthly active users were about 650,000, but by the end of June, they had increased about 20-fold to approximately 13 million. Daily token processing surged from 300 billion to 7 trillion, a scale exceeding the total processing volume of the independent OpenRouter (about 6 trillion).
In terms of revenue, OpenCode monetizes through two channels: token usage fees on its own inference infrastructure and a $10/month (approx. 1,637 yen) subscription called 'OpenCode Go.' Annualizing the token usage portion based on June data yields about $31–33 million (approx. 5.07–5.4 billion yen), but recent weekly data shows it jumping to about $38–40 million (approx. 6.22–6.55 billion yen). Combined with the annualized subscription portion of about $18 million (approx. 2.95 billion yen), the annualized revenue is on pace to reach about $56–58 million (approx. 9.17–9.5 billion yen). The subscription service has acquired about 160,000 users since its launch in March. Note that the inference service began between late September and early October last year, meaning it reached this level in about eight months.
The Block as a Catalyst: An 'Equal' Relationship with Anthropic
The inflection point of this growth curve was the friction with Anthropic that occurred in January 2026. At the beginning of the year, OpenCode's monthly active users were about 650,000, but in the first week of January, rumors began to circulate that Anthropic was trying to restrict people from using their Claude Code subscriptions on OpenCode. Many users were using Claude Code subscription credentials to access Claude models on OpenCode, and Anthropic began blocking requests containing the word 'OpenCode' in the system prompt to stop this. Jay stated, 'Since Anthropic is subsidizing the usage, I can understand the action itself,' but reflects that, as a result, this elevated OpenCode to the 'same playing field' as Claude Code and boosted the company's visibility at once. While many users were unhappy with this measure, the commotion spread the realization that OpenCode was not just 'one of many coding agents,' and people who hadn't used it before began to pay attention.
The program host analyzed this phenomenon by comparing it to Amazon's acquisition of Whole Foods leading to Instacart's rapid growth, noting that 'because it was blocked, people started taking OpenCode seriously for the first time.' Furthermore, according to a tweet by Tibo, the lead engineer of Codex, about 5% of Codex subscribers are choosing OpenCode as their main harness for Codex. Codex officially supported OpenCode shortly after the exchange with Anthropic.
The Mission of 'Delivering Magic to the World'
Jay says, 'Most people in the world have not yet experienced the magic of coding agents.' He positions this market as unprecedented, stating that a market for intelligence has not existed until now, and urges that everyone should 'think in a positive-sum spirit to grow the pie.' He positions the feeling of using a coding agent for the first time as magical, and an important moment in technological history that happens once in a generation. On the other hand, labs providing frontier models have very high token costs, making the experience economically difficult for many people around the world. OpenCode was designed with the goal of delivering that magical moment to as many people as possible.
When the product was first launched in June last year, the main usage was users trying their own Claude Code subscriptions on OpenCode. However, around August to September of the same year, the first open-source models like GLM, Kimi, and MiniMax began to appear. At the time, there was a sense that open-source models were about six months behind frontier models, but their emergence triggered a wave of users to OpenCode as the only place to try multiple models. And as open-weight models reached a quality sufficient for actual work, using them in combination with OpenCode became realistic.
The '$10' Coding Agent Capturing Emerging Markets
Another pillar supporting OpenCode's growth is its global adoption. Looking at the distribution of paid plan users, China accounts for 17% of the total, ranking first, followed by Indonesia at 4%, Brazil at 5%, and Vietnam, penetrating deeply into regions where the $200/month (approx. 32,738 yen) Claude Code subscription is very expensive. It is highly likely that having so much usage in China is unprecedented for a Y Combinator-backed company; in that country, many AI models are domestic, and it is thought that one reason for usage is that OpenCode provides the freedom to choose among them.
At the same time, it is seeing unexpected growth in the US as well. This subscription plan was originally designed for developers around the world, and the US market was not initially in mind, because Silicon Valley developers 'spend money on tokens like water.' However, it is now growing rapidly in the US as well, which is a sign of a 'change in atmosphere' where people are becoming 'a bit more conscious about tokens,' and users wanting to try specific models like GLM 5.2 are flowing into the subscription plan. Internal usage on a scale of thousands of people is spreading across many Fortune 500 companies, and cases where companies rush for formal adoption, saying 'the number of users is increasing, so please answer our security questionnaire,' are occurring frequently. Jay laughs, saying, 'When enterprises pester you to sign security agreements, you know you really have product-market fit,' as he is sometimes asked by corporate users, 'I don't know who you are, but many employees are using it, so please sign a contract.'
On the other hand, regarding why large US companies, whose token budgets are effectively unlimited, use OpenCode, Jay explains, 'Many companies start using it because they don't want to be locked into a specific model or harness.' Users are simply looking for more choices, and OpenCode provides the flexibility to switch to any model in the future as a 'neutral option.' Jay emphasizes that such adoption is the result of the product design, which was intentionally designed for everyone to use every day, resonating more than just cheap token prices. Note that in DMs and emails from companies, there is often a request to 'please don't make this public.'
The 'True' Model Popularity Shown by Data: DeepSeek's Potential and the Gemini Reversal
OpenCode provides a public dashboard (opencode.ai/data) that visualizes the usage status of a wide variety of models. The published data is based on usage on the 'OpenCode Go' subscription plan, which allows unlimited use of any open-source model for $10 per month, with daily token volume trends broken down in detail by model.
According to this, the most used model by token volume is DeepSeek Flash, followed by DeepSeek Pro, and then GLM 5.2. This data can also be viewed based on the number of unique users, where DeepSeek Flash has approximately 38,000 users, DeepSeek Pro has approximately 31,000, and GLM 5.2 has approximately 30,000, showing they are neck and neck. Jay points out, 'While there is a lot of talk on social media that GLM has overtaken DeepSeek, the actual data says otherwise.' GLM is on par with one DeepSeek model, but because DeepSeek has two models, there is still a difference in overall capability. The reason DeepSeek Flash is chosen is not just its low cost, but because it is significantly faster than services like Novus, giving developers a 'sense of working in real-time.' On the other hand, GLM 5.2 is highly rated for its strong front-end design, showing that users switch between models depending on the task.
What is particularly interesting is that when looking at the market share graph, DeepSeek's share drops temporarily when GLM appears, but recovers immediately afterward. Jay analyzes that because of DeepSeek Flash's low cost, a behavior is spreading where users switch to this cheaper model to complete the rest of their work when they approach their daily or weekly usage limits. This is a very different way of using tokens compared to the common Silicon Valley mindset of using them like water.
Furthermore, there is another symbolic turning point that appeared in OpenCode's data. In February of this year, for about four weeks, a reversal was observed for the first time in history where users on OpenCode used Gemini (Gemini 2.4 or 2.5 in the company's recognition) far more than Anthropic's models (the combined total of Sonnet and Opus at the time). Until then, even though open-source models were cheaper, the vast majority of usage was always occupied by cutting-edge frontier models, so Jay reflects, 'At this moment, I was convinced that we should launch a subscription product. I could determine that these models were sufficient for actual work.'
The Reality of Enterprise Adoption: Demands Beyond Procurement
After overcoming hurdles in procurement and management, companies have various requests. One is the voice saying, 'We want non-technical employees to use OpenCode,' and another is, 'We want to incorporate coding agents into our company's core product loop, so can we use it as a foundation for that?' Furthermore, there are inquiries from specific organizations within companies that do not need frontier models and want to manage access restrictions and token spending a bit more creatively. Among these, what Jay recalls as 'strange' is the request to 'visualize in detail who is doing what within the company,' an area they are considering whether to build as part of OpenCode.
The Structure Where Free Users Become the Catalyst for Enterprise Adoption
Another characteristic structure in OpenCode's growth is the flow where users who come in from the free tier turn into powerful advocates who promote adoption within Fortune 500 companies. This is not just about 'using it because it's cheap.' In addition to meeting the needs of companies that dislike vendor lock-in, it is also supported by a 'change in atmosphere' among large companies that have begun to be conscious of token costs. Even without heavily subsidizing token costs like large AI labs, they provide the first 'aha moment' in the free tier, lead them to a subscription from there, and eventually create 'whale' users who think it is worth investing more to rebuild their own processes. OpenCode can be said to be a rare entity that has established this 'customer acquisition cost through tokens' composition without relying excessively on subsidies.
Ramp's Advanced Use Case: OpenCode as a Slack Bot
The case of Ramp symbolizes how OpenCode is used in companies. Ramp published a blog post in December 2025, but prior to that, an internal team had independently built a Slack bot that runs OpenCode on the backend. Jay expresses his shock, saying, 'It was incredibly amazing. They did something we hadn't even done internally yet, and showed us a use case that points to the future.'
Here, it is necessary to understand OpenCode's architecture. OpenCode consists of two main parts: the UI that users see and interact with (the application part), and the 'agent loop' that actually calls the LLM to execute work, which the company calls the 'server.' This server can be separated from the UI and incorporated independently, and at Ramp, they took that server part and ran it behind a Slack bot.
Structural Changes in the AI Token Economy and Anthropic's Profitability
At this point, a fundamental change is occurring in the economic structure of AI companies. In traditional B2C and B2B companies, the majority of customer acquisition cost (CAC) was advertising expenses. But now, it is shifting to a model of acquiring users through the provision of tokens. At the same time, a structure is established where some large users (whales) pay high token fees, and even if there is a lot of churn among other users, it works if power users and experts convert the entire organization of a large company.
In this context, what should be noted is the revenue data for Anthropic mentioned on Dylan Patel's podcast. Anthropic is expected to achieve an astonishing figure of $50 billion in annualized revenue in the second quarter with a profit margin of about 70%, and since they were in the red and had significantly less profit the previous year, it can be said that they have truly 'crossed the chasm.' This structure overlaps with the direction OpenCode is heading, but the big difference is that OpenCode does not need to 'subsidize' token costs on a large scale like Anthropic. As Jay says, the free tier and subscription are equivalent to CAC where 'token costs are borne instead of advertising expenses,' and the amount that whales pay per token directly contributes to the margin.
The Monetization Barrier Crossed Without 'Subsidizing' Tokens
Anthropic and OpenAI adopt a model where they effectively subsidize those costs with subscriptions, getting users to cross that barrier and making a profit from usage-based billing from some heavy users. OpenCode provides an 'aha moment' in the free tier and raises it to the 'practical' stage with a $10/month subscription that leverages the low cost of open-source models. It is a structure where processing a large amount of tokens yields volume discounts on the inference infrastructure, which directly leads to margins on usage-based billing.
'Symbiosis' as the Largest Customer of Open Source Models
OpenCode is currently the largest token consumer for many open-source model providers. Jay describes this relationship as a 'symbiotic relationship,' stating, 'We are betting on making the entire open-source ecosystem work.' As the commoditization of models progresses, his argument is that a structure where an 'app layer' like OpenCode owns the touchpoint with users and encourages healthy competition between models brings more benefits to users than being locked into a specific vendor. In a situation of vendor lock-in, users cannot enjoy the benefits of competition, and there is a risk that vendor margins will increase. OpenCode's growth is a proxy indicator that the choices of available models have improved over time, and the growth in monthly active users is directly linked to the prosperity of the open-source model market.
Because the market is so huge, it is easy to imagine that each model lab will acquire its own niche specialized in specific areas or characteristics. In fact, DeepSeek has chosen the axis of 'cost performance' and has succeeded in excelling there. Jay says, 'As the labs themselves say, this is an unprecedented intelligence market, and everyone should approach it with a positive-sum spirit. If you think about it that way, OpenCode's growth means the birth of a huge customer for all labs.'
Also, due to the characteristic of being used around the world 24 hours a day without rest, the waves of GPU usage are leveled out, and maintaining a high utilization rate is a factor that allows OpenCode to have an advantage in inference costs. OpenCode rents GPUs and collaborates directly with inference-only providers and model labs, but because users are dispersed around the world, the peaks in the East and the peaks in the West complement each other, realizing stable 24-hour GPU cycles.
A Product Philosophy Aiming for an Open Default
At the foundation of OpenCode's product design is a clear strategy to become the "open default." The name "OpenCode" itself expresses this intent. Drawing on Jay's past experience with a similar project called "OpenNext," he learned that in markets dominated by major players, the rest of the market tends to coalesce around open alternatives. Capturing that position faster than anyone else was crucial. To declare support for over 70 models and providers at launch, they launched another open-source project called models.dev, building a database of models and providers that didn't exist at the time. This has now grown into what is likely the best dataset covering every model and provider in the world.
A Product Born from a Commitment to "Buying Coffee via Terminal"
Another foundation of OpenCode's product design is the founding team's strong commitment to "terminal UI." When Claude Code appeared around February 2025, Jay intuitively felt, "This is a fundamentally different experience from traditional autocomplete." Jay and Frank have been long-time users of Neovim and Vim, and they were not particularly drawn to products like Cursor. While they were impressed by Claude Code's functionality, they were not satisfied with the terminal experience itself. What they aimed for was a "modern terminal experience" that would feel as natural to core developers as other excellent terminal UIs like Neovim.
This sensibility is also reflected in their past projects. Co-founder Dax, along with a friend, developed a hobby project called "SSHterminal.shop," a complete storefront where one could order coffee from a terminal via SSH. Additionally, Jay and his team had experience incorporating terminal UIs into the core of their main product, SST. While these might seem like strange obsessions at first glance, they embodied Paul Graham's teaching to "make what developers want." Because they understood better than anyone that the open-source community they were deeply rooted in valued an excellent terminal experience above all else, their "eccentric tastes" became a powerful differentiator that instantly captured their initial core support base.
From 0 to $30 Million: A 16-Year-Old Corporation, 8 Months of Rapid Growth
Today's success did not appear "overnight." Jay first aspired to start a business around 2006-2007, during his sophomore year of college. Inspired by Paul Graham's essays, he applied to YC for the first time around the same period as Mixpanel and Airbnb. He incorporated in 2010, and with co-founder Frank, they have literally continued under the same corporate entity, challenging various waves from cloud and mobile to serverless. They applied to YC nine times starting in 2016, and after four interviews, they finally participated in 2021 with an idea for a serverless framework. Jay laughs, saying, "Normally, it would have been wiser to fold the company and get a job at a fast-growing startup," but the experience of making steady progress and learning the big picture of management—including marketing and positioning—has now culminated in OpenCode, which captures the market from consumers to enterprises in one go. Furthermore, following advice from Dalton, they thoroughly embraced "building in public," and their stance of making everything public created a culture where the community follows the company's journey as if watching a reality show.
2006: Starting a Business as a College Dropout and the Aspiration for YC
The story goes back to 2006-2007. Jay, a sophomore at the University of Waterloo, felt no attraction to employment and decided to start a business with his roommate, Frank. At 21, Jay thought he was "special" and named the company "Anomaly." The first thing he devoured were Paul Graham's essays. His first application to Y Combinator was during that period, and he stepped into Silicon Valley for his first interview. On the day of the interview, the person sitting next to him in the waiting room was none other than the founder of Airbnb. The interviewer was Paul Graham himself.
9 Applications, 4 Interviews, and Acceptance in 2021
The road from there was not smooth. Counting from 2016, they applied to YC nine times and made it to the interview stage four times. Although they challenged with different business ideas each time, these were all continued by the same corporate entity and the same founders, Jay and Frank. The turning point came in 2021. The idea with which they were accepted into YC was a "serverless platform." It was a concept to realize Heroku on AWS serverless and expand the market. The serverless framework they developed during this process became their first large-scale open-source project. After participating in YC, a third co-founder, Dax, joined.
Awakening to "Building in Public" and 10 Years of Dormancy
There were two decisive triggers for the culture of "building in public" to take root. One was the encouragement from YC partner Dalton Caldwell, who said, "Since you are an open-source company, you should operate in the public eye." The other was around 2022, when co-founder Dax argued, "Even though the code is all public, not talking about it publicly is a disadvantage to ourselves." Since then, they have continued to share information with surprising transparency, and that stance has brought the community a sense of "following the team and the company's journey as if watching a reality show."
Over these more than 10 years, they also experienced running a consumer-facing company and honed their skills in user acquisition and metrics management. They failed many times and even had periods where they ran out of money and lived at their parents' homes, but Jay reflects that "being a bit stubborn and always feeling positive progress" were the reasons they were able to keep going without giving up. Jay says, "Now that I can think about the entire customer journey consistently, I don't over-optimize for a specific segment, and it's much more fun."
Summary
At the end of the program, as a message to developers trying out OpenCode, Jay succinctly described its appeal: "It's the best place to try out new open-source models as soon as they come out." Also, MC Dalton shared his own experience of using OpenCode to submit a pull request to a large and complex codebase, stating that he was impressed by its high practicality even when combined with open-source models. He suggested that for those who have only used Claude Code or Cursor, it is well worth considering a transition to OpenCode.
#Startup #AI #YCombinator #OpenCode #JayV #GenerativeAI #CodingAgent #OpenSource #Anthropic #ClaudeCode #DeepSeek #GLM #UnitEconomics #EmergingMarkets #TerminalUI #Entrepreneur #Tenacity #TokenEconomy #BuildingInPublic #EnterpriseAdoption #ProductMarketFit #AnthropicRevenue #Ramp #SlackIntegration #ModelLab #Gemini #PaulGraham #Airbnb #Serverless #Neovim #DaltonCaldwell
