Opencode CEO: Blocked, 20X Growth in 6 Months, Building the Coding Agent for the World
In a Nutshell
OpenCode grew from 650K to 13M monthly users and 300B to 7T daily tokens in six months, reaching $38-40M annualized revenue with 160K subscribers. The company succeeded as an open-source interface to any model, surviving Anthropic's January block that backfired and boosted adoption, while dominating open-source model usage especially in emerging markets. Its 24-hour global usage cycle and marketplace approach—letting users switch between frontier and cheap open models—creates sustainable economics without heavy subsidization.
These notes were generated by AI and may contain inaccuracies.
Most people in the world still haven't experienced the magic of a coding agent. This is an unprecedented market because the market for intelligence has not existed before. The growth mentality should be positive sum. OpenCode sees significant traffic from developing countries including Indonesia at 4% and Brazil at 5%, along with Vietnam and similar markets where a $200 monthly Cloud Code subscription is prohibitively expensive.
Product-market fit is evident when enterprises proactively request security agreements to use the product, with companies stating they have users internally who want to use it and requesting the paperwork to formalize adoption.
OpenCode is an open-source alternative to Cloud Code that works with any model. The company went through YC in 2021. Current metrics include 4.6 million weekly active users, approaching Codex subscriber numbers. Monthly active users reached approximately 13 million by end of June, representing a 20x increase since the beginning of the year.
Token processing has grown from approximately 300 billion per day at the beginning of the year to 7 trillion tokens per day currently. For context, Open Router processes around 6 trillion tokens total.
Revenue from the subscription product and pay-per-token inference, extrapolated from June data, projects to $31-33 million annually. Using last week's data, the annualized projection reaches $38-40 million. The inference portion of the business launched end of September to early October of the previous year, achieving this scale in approximately 8 months.
The subscription product launched end of February to early March and has grown to 160,000 monthly subscribers, accounting for approximately $18 million in annualized revenue.
Codex officially supports OpenCode, meaning users can utilize their Codex subscription within OpenCode. Approximately 5% of Codex subscribers choose OpenCode as their primary interface for accessing the underlying API.
In the first week of January, when monthly active users stood at 650,000, Anthropic attempted to block users from using OpenCode with Claude Code subscriptions. The blocking mechanism rejected requests when the system prompt contained the literal word "open code." While this action made business sense from Anthropic's perspective to prevent subsidized usage, it inadvertently elevated OpenCode's visibility by placing it on equal footing with Claude Code.
The incident created widespread awareness as users who weren't previously using OpenCode took notice that Claude Code was taking action against it, making the product seem worth investigating beyond the dozens of other coding agents available.
The product premise centers on bringing the magical first experience with coding agents to as many people globally as possible. Frontier models charge high per-token rates that make this experience inaccessible for many users worldwide. OpenCode aims to provide this experience more affordably.
When OpenCode first launched in June of the previous year, usage was primarily driven by users wanting to use their Cloud Code subscription through OpenCode. By August-September, the first wave of open-source models emerged from companies including GLM, Kimi, and MiniMax. Initially perceived as approximately 6 months behind frontier models, these open-weight models eventually became good enough for real work, making OpenCode viable as an interface for multiple models.
Until February of the current year, frontier models dominated usage despite open-source alternatives being cheaper. During a 4-week period, Gemini 2.5 usage exceeded combined Sonnet and Opus usage for the first time in the company's data. This shift prompted the launch of the subscription product, as it indicated these models had become viable for real work.
OpenCode publishes usage data at opencode.ai/data. The data specifically covers OpenCode Go, the $10 subscription plan for using open-source models. Token volume breakdown shows DeepSeek Flash as the most heavily used model, followed by other DeepSeek variants and GLM. Despite Twitter discussions suggesting GLM was overtaking DeepSeek, the actual usage data shows DeepSeek maintaining stronger position.
By unique users, DeepSeek Flash leads with approximately 38,000 users, DeepSeek Pro has 31,000 users, and GLM 5.2 has close to 30,000 users. DeepSeek Flash's low cost enables users to extend their coding agent usage when approaching daily or weekly limits by switching to cheaper models.
China represents 17% of usage on the OpenCode Go plan, making OpenCode likely the only YC company with meaningful usage in China. Chinese users prefer using models built in their country, and OpenCode provides this choice. The US market has grown significantly despite initial assumptions that American users would continue paying premium prices. Indonesia contributes 4% of traffic, Brazil contributes 5%, with strong adoption in Vietnam and other developing markets.
Large US companies with unlimited token budgets use OpenCode to avoid vendor lock-in to specific models or harnesses. Dozens of top Fortune 500 companies have significant usage footprints. These enterprises often discover the product through internal usage before formal procurement processes begin. Companies reach out requesting security questionnaires because employees are already using the product internally.
Enterprise requests focus on expanding usage beyond developers to non-technical users and integrating coding agents into product core loops. Additional requests involve token management capabilities, including limiting frontier model access for certain departments and creating visibility into usage patterns across the organization.
Ramp built a Slack bot using OpenCode's agent loop (the server component) running behind the scenes. This demonstrated embedding the agent functionality separately from the UI, showing a future use case the company had not yet implemented internally.
Unlike traditional B2C or B2B companies that acquire customers through advertising spend, AI companies face token-based customer acquisition costs. The free tier serves as the mechanism for users to experience the "magic moment" with coding agents. The subscription plan enables real work with cheaper open-source models. This two-tier approach allows users to cross the expensive chasm of understanding AI systems before potentially becoming high-value customers or converting their organizations.
Anthropic achieved profitability with 70% margins on a $50 billion annualized revenue run rate in Q2, having crossed this chasm through subsidization. OpenCode's approach differs by not needing to subsidize usage at the same scale due to the economics of open-source models.
Paying per token allows volume discounts based on the amount of tokens served. This effectively turns into margins, while subsidizing or offering a free tier means eating the cost. As more whales sign up, they pay per token which directly feeds into margins.
The question of where value will accrue - whether frontier model labs will capture everything or whether it will go the other way - remains central. OpenCode owns the relationship with end users and makes it easy for them to pick and choose models. This creates a marketplace dynamic where users select models based on different attributes and cost characteristics.
The approach showcases model diversity, creating an environment where labs are aware of each other and competition benefits consumers. Being locked into a specific vendor prevents benefiting from competition and may help that vendor's margins instead.
Growth serves as a proxy for the fact that choices have only improved over time. Every bump in monthly active users can be tracked back to some bump in the open source model market. The strategy involves betting the field rather than picking a winner among model labs.
Models are becoming commoditized utilities. The market is so large that model labs are picking off niches and chunks, specializing for specific areas or characteristics. DeepSeek has picked the cost part of the quality-cost-performance axes, demonstrating this specialization pattern.
Believing the sales pitch of the labs - that this is an unprecedented market for intelligence - leads to a positive-sum grow-the-pie mentality. Labs should want OpenCode to grow and succeed because it becomes a huge customer for all of them.
OpenCode is the largest customer for most open-source models in terms of token volume. This creates a symbiotic relationship where both parties need each other. The other half involves the open-source ecosystem as a whole, making the entire thing work.
GPUs are rented and work is done with providers offering straight inference and model labs themselves. Due to global usage, peaks and troughs over the course of a day for GPU utilization are reasonably stable. When the East is working, the West may be asleep, but when the West is working, the East is active, creating a stable 24-hour GPU cycle that allows for pretty good utilization and helps unit economics.
The stable 24-hour GPU cycle provides a competitive advantage when compared to counterparts that may just serve one part of the world, enabling more efficient model operation.
The name OpenCode literally comes from the positioning strategy. A previous project was called OpenNext, based on the idea that when there are dominant players in the market, the rest coalesces around an open alternative. Picking that position is valuable because it's very hard for somebody else to displace you, and if it's open, the goal should be to become the default as quickly as possible.
At launch, the claim was made to support 70 plus models and providers. To make this happen, a separate open source project called models.dev was created that built up an entire database that didn't exist at the time. This has become probably the best data set of all the models and providers out there in the world.
The first pivotal moment occurred when the previous product had just hit break-even around February 2025. Claude Code came out and was recognized as fundamentally different from using autocomplete with AI. It was something that made sense as a core developer tool. Being Neovim/Vim users at the time, Cursor didn't necessarily resonate as much. The realization was that Claude Code wasn't up to the standard of terminal UIs like Neovim, leading to the desire to build something that felt like a modern terminal experience.
A terminal UI was built for buying coffee online as a fun pet project. The idea was that as a hardcore developer in the terminal all day, you shouldn't be bothered to open up a browser. Users could go to SSHterminal.shop and order coffee over SSH.
Building things that developers want themselves, even if they seem goofy to others or VCs would dismiss them as horrible businesses, has a tendency to pull in interesting directions. Having eccentric tastes isn't always going to lead to something interesting, but it's kind of how outlying things get created. Being embedded in the open source community and surrounded by people who looked up to products that were really good in the terminal meant that building one would resonate with them almost right away.
The company's been around for longer than 6 months. The founders went through YC over 5 years ago, but the story starts way before that. The legal entity was incorporated in 2010, making it a 16-year-old legal entity. It took 16 years since starting the company to have a truly runaway success product.
The story began in second year university after a co-op term at Waterloo. The founder came back and, being very naive, found the smartest people available - Frank, the college roommate. The two started a company, with the 21-year-old founder picking the name Anomaly because he felt special. This was in 2006-2007.
The first YC application was probably around that time, with the first interview bringing the founder to Silicon Valley. This was the Mixpanel batch, with the interview happening literally after Mixpanel, right after Airbnb's batch. One of the Airbnb founders was hanging out in the room while waiting to be interviewed, and this was with PG back then. It took more than a decade to get into YC.
During those years, it wasn't different startups but literally the same startup - the same legal entity all those years. The founders did nine applications from 2016 up to 2021 when they got accepted, with four interviews. All were different ideas but the same legal entity and same founders.
After doing YC, the third co-founder started joining, making it three founders for the last four years.
The idea applied to YC with in 2021 was building a serverless platform - like Heroku but for AWS and serverless. The goal was to do a better job in that space and maybe grow the market. A serverless framework was built as the first big open source project and first move into building open source products and building in public.
A little bit of building in public came from Dalton pushing the idea that things should be done in public because it's an open source company. Around 2022, it dawned that all code is public and work happens basically in public, so not talking about it publicly would be doing a disservice to both the company and product. Ever since then, it's been part of the identity.
For a lot of the community following on Twitter, it feels like watching a reality TV show of the team, the company, and the journey they're on.
Even though it sounds like a lightning in the bottle moment, the reality is there's been grinding for a good 10 years without giving up. All the windy destinations that ended up being dead ends actually taught different things. Running a consumer company meant getting really good at consumer acquisition and tracking numbers, which plays into the current level of detail for OpenCode.
With OpenCode, the entire market can be addressed - consumers, individual users, small teams, the open source community, mid-market, all the way to enterprise. In past iterations of different products, one of each was probably done. Now it feels like getting to do all of them together in one product, thinking about the customer journey through the entire thing without hyper-optimizing for specific parts.
What got through the difficult periods was partly being a little stubborn. There was something in the back of the mind that felt like different things were being learned along the way - not just building a product, but doing marketing, understanding positioning, doing the whole thing. That journey felt like progress and positive progress. Part of that was being fortunate to have the ability to do that, basically living with parents for a bunch of time and running out of money. It was a combination of being a little thick-headed and seeing positive progress.
For folks who have only used Claude Code or CodeX or Cursor, when hearing about a new model that comes out, especially an open-source one, OpenCode is really good for trying it out. Users can go in, look at the model picker, and see GLM 5.2 or whatever the new open-source model is up there to pick and start using right away.
The story hits on classic startup wisdom: pursuing interests, having eccentric tastes, living in the future a little bit. The fact that the founders kept building their taste and building things over a long period of time, then when lightning strikes, they're actually in a position to capture it - that's the thing that doesn't get mentioned. To catch lightning in a bottle, the bottle actually has to be positioned correctly, be ready for it, and know what to do with it.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
